I—a
and
y{n) = Y a^-^ = a ^ ^ — % = ^ , N
^
h{n,k)u(k).
(16.11a)
We can rewrite (16.1 la) as
71 CHAPTER 1:
y{^) ^
^
h{n^k)u{k)-\- ^
k=-oo
Mathematical Descriptions of Systems
h{n^k)u{k)
k=ko
^y{ko-l)^f^h{n,k)u{k).
(16.11b)
k=ko
We say that the discrete-time system described by (16.9) is at rest 3tk = ko ^Z if u{k) =Ofork> ko, implies that y{k) =Ofork> ko. Accordingly, if system (16.9) is known to be at rest at ^ = ^o» we have y{^) = ^
h{n^k)u{k).
k=kQ
Furthermore, if system (16.9) is known to be causal and at rest at ^ = ^o» its inputoutput description assumes the form [in view of (16.1 lb)] y{^) = ^
(16.12)
h{n^k)u{k).
k=kQ
Next, turning to linear, discrete-time, MIMO systems, we can generalize (16.9) to y{n)= where y \Z ^RP
,u\Z
H{n,k)
X
H{n,k)u{k)
(16.13)
^W^.dind 'hii{n,k) h2i{n,k)
hi2{n,k) h22{n,k)
hpi{n,k)
hp2{n,k)
himin.k)h2m{n,k)
(16.14)
where hij{n^k) represents the system response at time n of the i\h component of y due to a discrete-time impulse 5 applied at time k at the jth component of u, while the inputs at all other components of u are being held zero. The matrix H is called the discrete-time unit impulse response matrix of the system. Similarly, it follows that the system (16.13) is causal if and only if H{n^ k) =0
for all k and all n < k,
and that the input-output description of linear, discrete-time, causal systems is given by n
y{n)=
^
H{n,k)u(k).
(16.15)
A discrete-time system described by (16.15) is said to be at rest at k = ko ^Z if u{k) = 0 for ^ > ^0 implies that y{k) = 0 for ^ > ^o- Accordingly, if system (16.13) is known to be at rest at ^ = ^o, we have y{n)=
J^H{n,k)u{k). k=ko
(16.16)
72 Linear Systems
Moreover, if a linear discrete-time system that is at rest at ko is known to be causal, then its input-output description reduces to n
y(n) = ^
H(n, k)u(k).
(16.17)
k=ko
Finally, as in (16.10), it is easily shown that the unit impulse response H(n, k) of a linear, time-invariant, discrete-time MIMO system depends only on the difference of n and k, i.e., by a slight abuse of notation we can write H{n, k) = H(n - k,0) = H(n - k)
(16.18)
for all n and k. Accordingly, linear, time-invariant, causal, discrete-time MIMO systems that are at rest at k = ko are described by equations of the form n
y(n) = ^H(n-
k)u(k),
(16.19)
We conclude by supposing that the system on hand is described by the state and output equations (15.3a) and (15.3b) under the assumption that x(ko) = 0, i.e., the system is at rest at ^ = ^o- Then, according to Eqs. (15.17) and (15.18), we obtain
C(«)0(n, k + \)B{k), Hin, k) =
n> k, n = k, n < k.
(16.20)
Furthermore, for the time-invariant case we obtain CA^'-^^^^^B, H{n -k) =
\D,
0,
n > k, n = k, n < k.
(16.21)
C. The Dirac Delta Distribution For any linear time-invariant operator P from C(R, R) to itself, we say that P admits an integral representation if there exists an integrable function (in the Riemann or Lebesgue sense), gp : R^ R, such that for any / G C(R, R), r 00
{Pf){x)
= ( / * gp){x) ^
f(T)gpix - T) dr. J -co
We call gp a kernel of the integral representation of P. For the identity operator / [defined by If = f for any / E C(R, R)] an integral representation for which gp is a function in the usual sense does not exist (see, e.g., Z. Szmydt, Fourier Transformation and Linear Differential Equations, D. Reidel Publishing Company, Boston, 1977). However, there exists a sequence of functions {(/)„} such that for any / G C(R, R), (If)(x)
= fix)
= l i m ( / * c/>,)(x).
(16.22)
73
To prove (16.22) we will make use of functions {(pn) given by
CHAPTER 1:
«(1 -- n\x\).
if IJCI <
0,
if \x\ > -, n
0«W = <
-
Mathematical Descriptions of Systems
n
n = 1, 2, 3, A plot of (/)„ is depicted in Fig. 1.16. We first establish the following useful property of {(/)„}.
FIGURE 1.16 LEMMA 16.1. L e t / be a continuous real-valued function defined on R and let (/>„ be defined as above (Fig. 1.16). Then for any a E R, lim
f{T)(f>n{a -T)dT
= f(a).
(16.23)
Proof, It is easy to verify that for n = 1, 2 , . . . , (/>„(T) dr =
4>n{T) dr = 1. J-l/n
Then (f)n{a — T) dr
(pnici^ ~ T) dr a-l/n \ln
(16.24)
(j)n{T)dT = 1.
We first assume t h a t / is a nonnegative function. By the continuity of/ we may suppose that/ assumes a maximum value f(bn) and a minimum value /(c„), where bn and Cn E \a ~ -,a + -\. Then I n n\ r a+l/n
(l>n(a - T)f{T) dr = 3
(f)n{a - T)f(T) dr Ja-l/n
r a+l/n
f(bn)
Ma
-T)dT
= f{bn),
(16.25)
Ja-\ln
where we have used (16.24). In a similar manner we can show that ct>n{a-T)f{T)dT^
f{Cn).
(16.26)
1 1 a - -,a + - , it follows that lim„_^oo bn = lim„^oo Cn = a. Thus, n n (16.25) and (16.26) together with the continuity of/ imply (16.23).
Since bn and Cn
74 Linear Systems
To remove the assumption that / is a nonnegative function, we recall that every continuous function / can be written as the sum of two nonnegative functions, / = /+—/_, where
.,
J/W, if/(x)>0,
^^^"[0,
if/(x)<0,
[-/(x),
if/(x)<0.
Then lim / f{T)(l)n{a-T)dT = lim /
/+(T)0„(a —T)(iT—lim /
/_(T)0„(a —T)(iT
= U{a)-U{a)=f{a). • The above result, when applied to (16.22), now allows us to J^^n^ a generalized function 5 (also called a distribution) as the kernel of di formal or symbolic integral representation of the identity operator/, i.e., f{x) = lim r = /
f{T)Ux-T)dT
(16.27)
/(T)5(JC-T)JT
(16.28)
= /*5(jc). (16.29) /f is emphasized that expression (16.28) is not an integral at all (in the Riemann or Lebesgue sense), but only a symbolic representation. The generalized function d is called the unit impulse or the Dirac delta distribution. In applications we frequently encounter functions / G C{R^,R). If we extend / to be defined on all ofRby letting f{x) = 0 for x < 0, then (16.23) becomes lim /
f{T)^n{ci-T)dT
= f{a)
(16.30)
for any a > 0, where we have used the fact that in the proof of Lemma 16.1 we need / to be continuous only in a neighborhood of a. Therefore, for / G C{R^,R), (16.27) to (16.30) yield lim rf{T)Ut-^)dT^ rf{T)5{t-T)dT = f{t) (16.31) n^^Jo Jo for any t > 0. Since the 0^ are even functions, we have 0^(f — T) = 0^(T — f), which allows for the representation d{t — T) = d{T— t). We obtain from (16.31) that lim rf{T)U^-t)dT^ rf{T)5{T-t)dT n^^Jo Jo for any t > 0. Changing the variable T^ = T — f, we obtain lim r f{T'+t)U^')dT'^
r f{T' +
= f{t)
t)5{T')dT'=f{t)
for any t > 0. Taking the limit f ^ 0+, we obtain lim r f{T'+ t)UT')dT' n^^Jo-
^ r f{T')5{T')dT' Jo-
= f{Q),
(16.32)
where, as in (16.24), J^- f{T')5{T')dT' is not an integral, but a symbolic representation 0f\imn^ooSQf{T'+t)(^n{^')dT'. Now let s denote a complex variable. If in (16.31) to (16.32) we let / ( T ) = e~^^ ^ T > 0, then we obtain the Laplace transform lim / e-'^(l)n{T)dT= I e-'^5{T)dT=l. n^^ JoJoSymbolically we denote (16.33) by ^ ( 5 ) = 1,
(16.33)
(16.34)
and we say that the Laplace transform of the unit impulse function or the Dirac delta distribution is equal to one. Next, we point out another important property of 5. Consider a (time-invariant) operator P and assume that P admits an integral representation with kernel gp. If in (16.31) we let / = gp, we have \im{P^n){t)=gp{t),
(16.35)
and we write this (symbolically) as PS = gp.
(16.36)
This shows that the impulse response of a linear, time-invariant, continuous-time system with integral representation is equal to the kernel of the integral representation of the system. Symbolically this is depicted in Fig. 1.17. Next, for any linear time-varying operator P from C{R^R) to itself, we say that P admits an integral representation if there exists an integrable function (in the Riemann or Lebesgue sense), gp : RxR^ R, such that for any / G C{R,R),
{Pf){ri) =
iyit)gp{ri,t)dt.
(16.37)
Again, we call gp a kernel of the integral representation of P. It turns out that the impulse response of a linear, time-varying, continuous-time system with integral
9)1
'n
• • • • 5 FIGURE 1.17
p
• • • • p
9n
• • • • 9p
oo
75 CHAPTER 1:
Mathematical Descriptions of Systems
76 Linear Systems
representation is again equal to the kernel of the integral representation of the sys^^^- To see this, we first observe that if /z G C{R X R, R), and if in Lemma 16.1 we replace / G C{R, R) by h, then all the ensuing relationships still hold, with obvious modifications. In particular, as in (16.27), we have for all t ^ R, lim
h{t, T)(l>n(V -T)dT
= I
h{t, 7)8(7] - T) dj = h(t, 7]). (16.38)
Also, as in (16.31), we have lim h(t, T)(j)n{r] -T)dT «^°°Jo
= Jo
h(t, T)8(r] -T)dT
= hit, r/)
(16.39)
for 7]>0. Now let hit, T) = grit, r). Then (16.38) yields r 00
lim
^00
gpit, r)cl)niv -r)dr
=
gp(t, 7)8(7] -7)d7
= grit, T]),
(16.40)
which establishes our assertion. The common interpretation of (16.40) is that gp(t, 17) represents the response of the system at time t due to an impulse applied at time r/. In establishing the results of the present subsection, we have made use of a specific sequence of functions {(/>„} given by
n(\ -- n\x\),
if Ixl < n ,
0,
if \x\ > n
^n(x) = {
n=
\,2,,...
We wish to emphasize that there are many other functions that could have served this purpose. An example of a commonly used sequence of functions employed in the development of the Dirac delta distribution is given by {ipn}, where ^n(t) = \ In'
I 0,
'^' " ""'
\t\ > n,
n = \,2, Another example of a sequence of functions that can be utilized for this is given by {r/^}, where f n^te'""^, Vn(t) = \ 0,
0 < ^ < 00, n=
1,2,..,.
elsewhere.
D. Linear Continuous-Time Systems We let P denote a linear time-varying operator from C(R, R^) = Uto C(R, RP) = Y and we assume that P admits an integral representation given by y(t) = (Pu)(t) = i
Hp(t,7)u(7)d7,
(16.41)
where Hp : R X R^ j^pxm^ u ^ U, and j G F, and where Hp is assumed to be integrable. This means that each element of Hp, hp.. : RX R-^ Ris integrable (in the Riemann or Lebesgue sense).
Now let yi and j2 denote the response of system (16.41) corresponding to the input ui and U2, respectively, let ai and 0^2 be real scalars, and let y denote the response of system (16.41) corresponding to the input aiui + 0:2^2 = u. Then y = P(u) — P(aiU\ + ^2^2) — — ai
\
Hp(t, T)[aiUi(T) +
Hp{t, T)U\(T) dr -\- a2 \
J—00
Hp(t,
«2W2(T)]
T)U2(T)
dr
dr
J—cc
= aiP(u\) -\- a2P{u2) = cxiyi + a2y2y
(16.42)
which shows that system (16.41) is indeed a linear system in the sense defined in (16.4). Next, we let all components of U(T) in (16.41) be zero, except for the jth component. Then the /th component of y(t) in (16.41) assumes the form yi(t)
=
(16.43)
hp.j(t,T)Uj(T)dT. J —00
According to the results of the previous subsection [see Eq. (16.40)], hp.j(t, r) denotes the response of the /th component of the output of system (16.41), measured at time t, due to an impulse applied to thejth component of the input of system (16.41), applied at time r, while all the remaining components of the input are zero. Therefore, we call Hp(t, r) = [hp.j(t, r)] the impulse response matrix of system (16.41). Now suppose that it is known that system (16.41) is causal. Then its output will be identically zero before an input is applied. It follows that system (16.41) is causal if and only if Hp(t,
T)
= 0
for all r and for all t < r.
Therefore, when system (16.41) is causal, we have in fact that y(t)
=
Hp{t, T)U{T)
(16.44)
dr.
We can rewrite (16.44) as y(t) =
Hp{t, T)U(T) dr +
= y(to) +
Hp(t,
T)U{T)
Hp(t,
T)U(T)
dr
(16.45)
dr.
We say that the continuous-time system (16.41) is at rest at t = to if u(t) = 0 for r ^ to implies that y(t) = 0 for ^ > ^o- Note that our problem formulation mandates that the system be at rest at to = -^. Also, note that if a system (16.41) is known to be causal and to be at rest at ^ = ^o^ then according to (16.45) we have y(t)
Hp(t,
T)U(T)
dr.
(16.46)
J to
Next, suppose that it is known that the system (16.41) is time-invariant. This means that if in (16.43) hp..(t, r) is the response yt at time t due to an impulse applied at time r at thejth component of the input [i.e., UJ(T) = 8(t)], with all other input components set to zero, then a - r time shift in the input [i.e., Uj(t - r) = 8{t - r)] will result in a corresponding - r time shift in the response, resulting in hp. .(t-T,0).
77 CHAPTER 1:
Mathematical Descriptions of Systems
78 Linear Systems
Since this argument holds for all t,TGR and for all / = 1 , . . . , p, and j = 1 , . . . , m, we have Hp(t, r) = Hp(t - r, 0). If we define (using a sHght abuse of notation) Hp{t - T,G) = Hp{t - T), then (16.41) assumes the form y(t) =
Hp(t -
T)U(T)
dr.
(16.47)
J —cc
Note that (16.47) is consistent with the definition of the integral representation of a linear time-invariant operator introduced in the previous subsection. The right-hand side of (16.47) is the familiar convolution integral of Hp and u and is written more compactly as y(t) = (Hp * u)(t).
(16.48)
We note that since Hp(t - r) represents responses at time t due to impulse inputs applied at time r, then Hp(t) represents responses at time t due to impulse function inputs applied at r = 0. Therefore, a linear time-invariant system (16.47) is causal if and only if Hp(t) = 0 for all t < 0. If it is known that the linear time-invariant system (16.47) is causal and is at rest at ^0, then we have rt
Hp(t-T)u(T)dT.
y(t) =
(16.49)
to
In this case it is customary to choose, without loss of generality, to = 0. We thus have y(t) = f Hp(t - T)U(T) dr, t > 0. (16.50) Jo If we take the Laplace transform of both sides of (16.50), provided it exists, we obtain y{s) = Hp(s)uis\
(16.51)
where y(s) = [yi{s\ .. .,ypJis)f,Hp(s) = [hp.j(s)l md u(s) = [ui(s\ ..., Um(s)]\ where the yi(s), Uj(s), and hp..(s) denote the Laplace transforms of yi{t), Uj(t), and hpij(t), respectively. Consistent with Eq. (16.34), we note that Hp(s) represents the Laplace transform of the impulse response matrix Hp(t). We call Hp(s) a transfer function matrix. Now suppose that the input-output relation of a system is specified by the state and output equations (14.3a), (14.3b), repeated here as X = A(t)x + B(t)u
(16.52a)
y = C(t)x + D(t)u.
(16.52b)
If we assume that x(to) = 0 so that the system is at rest at to = 0, we obtain for the response of this system, y(t) = [ C{t)^{t,T)B{T)u{T)dT = [ {C{t)^{t, JtQ
T)B(T)
+ D{t)u{t)
+ D(t)8(t -
T)]U(T)
(16.53) dr,
(16.54)
where in (16.54) we have made use of the interpretation of 8 given in Subsection C. Comparing (16.54) with (16.46), we conclude that the impulse response matrix for system (16.52a), (16.52b) is given by M (tr^-l
^^^)*(^^ ^)^(^) + ^WS(r - T),
t^
T,
(16.55)
Finally, for time-invariant systems described by the state and output equations (14.7a), (14.7b), repeated here as (16.56a) (16.56b)
X = Ax + Bu y = Cx -\- Du, we obtain for the impulse response matrix the expression Hp(t
r) = f Ce'^^'-^^B + D8(t - r), 0,
r > T, t < T,
(16.57)
or, as is more commonly written. Hp(t) =
Ce^'B + D8(t\ 0,
t^ 0, t<0.
(16.58)
We will pursue the topics of this section further in Chapter 2.
1.17 SUMMARY In this chapter we addressed mathematical descriptions of systems. First, initialvalue problems determined by systems of first-order nonlinear ordinary differential equations were introduced. Conditions for existence and uniqueness of solutions, and approaches to determine such solutions, were established. The main reason for considering nonlinear mathematical descriptions of systems in a book on linear systems is that the origins of most linear systems are nonlinear systems. Specifically, linear systems are frequently obtained from nonlinear systems by a linearization process, or are the result of modeling where the nonlinear effects of a physical process have been suppressed or neglected. Accordingly, the validity of linear mathematical descriptions must always be interpreted in the context of the nonlinear systems they approximate. The linearization of nonlinear systems along a given solution (and a given input) was discussed in Section 1.11. The solutions of linear (time-varying) state equations were obtained in Section 1.13 using the method of successive approximations. The response of a linear continuous time system to an input, given initial states, was derived in Section 1.14. The development of the theory of discrete-time systems parallels that of continuous-time systems and was addressed in Section 1.15. State-space representations provide detailed descriptions of the internal behavior of a system, while input-output descriptions of systems emphasize external behavior and how a system interacts with its environment. Input-output descriptions of linear systems, addressed in Section 1.16, involve the convolution integral for continuous-time systems and the convolution sum for discrete-time systems.
79 CHAPTER 1:
Mathematical Descriptions of Systems
80 Linear Systems
In the next chapter, both state-space descriptions and input-output descriptions ^f systems are revisited, and their dynamic behavior is studied in detail.
1.18 NOTES Standard references on linear algebra include Birkhoff and MacLane [2], Halmos [9], and Gantmacher [8]. For more recent texts on this subject, refer, e.g., to Strang 122] and Michel and Herget [12]. Excellent sources on analysis at the elementary level include Apostol [1], Rudin [17], and Taylor [24]. For treatments at an intermediate level, consult, e.g., Royden [16], Taylor [23], Naylor and Sell [15], and Michel and Herget [12]. For a classic reference on ordinary differential equations, see Coddington and Levinson [6]. Other excellent sources include Brauer and Nohel [3], Hartman [10], and Simmons [21]. Our treatment of ordinary differential equations in this chapter was greatly influenced by Coddington and Levinson [6] and Miller and Michel [14]. An original standard reference on linear systems is Zadeh and Desoer [25]. Of the many excellent texts on this subject, the reader may want to refer to Brockett [4], Kailath [11], and Chen [5]. For more recent texts on linear systems, consult, e.g., Rugh [18] and DeCarlo [7]. In Section 16 we showed that continuous-time finite-dimensional linear systems described by state equations have an input-output description given by integral equations. For a general and comprehensive treatment of the integral representation of linear systems, refer to Sandberg [19], [20].
1.19 REFERENCES 1. T. M. Apostol, Mathematical Analysis, Addison-Wesley, Reading, MA, 1974. 2. G. Birkhoff and S. MacLane, A Survey of Modern Algebra, Macmillan, New York, 1965. 3. F. Brauer and J. A. Nohel, Qualitative Theory of Ordinary Differential Equations, Benjamin, New York, 1969. 4. R. W. Brockett, Finite Dimensional Linear Systems, Wiley, New York, 1970. 5. C. T. Chen, Linear System Theory and Design, Holt, Rinehart and Winston, New York, 1984. 6. E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations, McGrawHill, New York, 1955. 7. R. A. DeCarlo, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1989. 8. F. R. Gantmacher, Theory of Matrices, Vols. I, II, Chelsea, New York, 1959. 9. R R. Halmos, Finite Dimensional Vector Spaces, D. Van Nostrand, Princeton, NJ, 1958. 10. P. Hartman, Ordinary Differential Equations, Wiley, New York, 1964. 11. T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980. 12. A. N. Michel and C. J. Herget, Applied Algebra and Functional Analysis, Dover, New York, 1993. 13. A. N. Michel and K. Wang, Qualitative Theory of Dynamical Systems, Marcel Dekker, New York, 1995.
14. R. K. Miller and A. N. Michel, Ordinary Differential Equations, Academic Press, New York, 1982. 15. A. W. Nay lor and G. R. Sell, Linear Operator Theory in Engineering and Science, Holt, Rinehart and Winston, New York, 1971. 16. H. L. Royden, Real Analysis, Macmillan, New York, 1965. 17. W. Rudin, Principles of Mathematical Analysis, McGraw-Hill, New York, 1976. 18. W. J. Rugh, Linear System Theory, Second Edition, Prentice-Hall, Englewood Cliffs, NJ, 1996, 19. I. W. Sandberg, "Linear maps and impulse responses," IEEE Transactions on Circuits and Systems, Vol. 35, No. 2, pp. 201-206, 1988. 20. I. W. Sandberg, "Integral representations for linear maps," IEEE Transactions on Circuits and Systems, Vol. 35, No. 5, pp. 536-544, 1988. 21. G. F. Simmons, Differential Equations, McGraw-Hill, New York, 1972. 22. G. Strang, Linear Algebra and Its Applications, Academic Press, New York, 1980. 23. A. E. Taylor, Introduction to Functional Analysis, Wiley, New York, 1958. 24. A. E. Taylor, Advanced Calculus, Blaisdell, New York, 1965. 25. L. A. Zadeh and C. A. Desoer, Linear System Theory—The State Space Approach, McGraw-Hill, New York, 1963.
1.20 EXERCISES 1.1. (Hamiltonian dynamical systems) Conservative dynamical systems, also called Hamiltonian dynamical systems, are those systems that contain no energy-dissipating elements. Such systems with n degrees of freedom can be characterized by means of a Hamiltonian function H{p, q), where q^ = {q\,..., qn) denotes n generalized position coordinates and p^ = {p\,..., Pn) denotes n generalized momentum coordinates. We assume that H{p, ^) is of the form H{p,q) = T(q,q) + W(qX
(20.1)
where T denotes the kinetic energy and W denotes the potential energy of the system. These energy terms are obtained from the path-independent line integrals T(q,q)=\
p(q,^fd^=\ JO
^pt(q,Od^i
(20.2)
Jo / ^ i
(20.3) where fj = 1 , . . . , n, denote generalized potential forces. For the integral (20.2) to be path-independent, it is necessary and sufficient that dpi{q,q) _ dqj
dpj{q,q) dqt
i,j=l,...,
n.
(20.4)
A similar statement can be made about Eq. (20.3). Conservative dynamical systems are described by the system of 2n ordinary differential equations qi = -z—{p,q), "^P^
I =
Pi = -—-(p,q), oqt
I =
l,...,n, (20.5) h...,n.
81 CHAPTER 1:
Mathematical Descriptions of Systems
82 Linear Systems
Note that if we compute the derivative of H(p, q) with respect to t for (20.5) [i.e., along the solutions qi{t), Pi{t\ i = I,.. .,n], then we obtain, by the chain rule. -^(p(t\q(t)) i=l
^
dH ^
JH^
,
^
OH ^
JH ^
,
^
^ - S ^^^' '^-^'' '^ ^ S ^^^' 'W' '^ ^ '• In other words, in a conservative system (20.5) the Hamiltonian, i.e., the total energy, will be constant along the solutions (20.5). This constant is determined by the initial data (p(0),^(0)). (a) In Fig. 1.18, Mi and M2 denote point masses, ^1, K2, K denote spring constants, and xi, X2 denote the displacements of the masses M\ and M2. Use the Hamiltonian formulation of dynamical systems described above to derive a system of first-order ordinary differential equations that characterize this system. Verify your answer by using Newton's second law of motion to derive the same system of equations. Assume that x\{0\ x\{0), ^2(0), i:2(0) are given.
A
A
K
1 h^A/W-
M,
Ko
^2 —VW
/77777777777V7777777777777777777777777A
i
FIGURE 1.18 Example of a conservative dynamical system (b) In Fig. 1.19, a point mass m is moving in a circular path about the axis of rotation normal to a constant gravitationalfield(this is called the simple pendulum problem). Here / is the radius of the circular path, g is the gravitational acceleration, and x denotes the angle of deflection measured from the vertical. Use the Hamiltonian formulation of dynamical systems described above to derive a system of first-order ordinary differential equations that characterize this system. Verify your answer by
FIGURE 1.19 Simple pendulum
using Newton's second law of motion to derive the same system of equations. Assume that x(0) and i:(0) are given, (c) Determine a system of first-order ordinary differential equations that characterizes the two-link pendulum depicted in Fig. 1.20. Assume that ^i(O), ^2(0), ^i(O), and ^2(0) are given.
y////////y
FIGURE 1.20 Two-link pendulum
1.2. (Lagrange's equation) If a dynamical system contains elements that dissipate energy, such as viscous friction elements in mechanical systems and resistors in electric circuits, then we can use Lagrange's equation to describe such systems. (In the following, we use some of the same notation used in Exercise 1.1.) For a system with n degrees of freedom, this equation is given by d idL
\
dL
..^dD
/ = 1 , . . . , n,
(20.6)
where q^ = {q\, ••-,qn) denotes the generalized position vector. The function L{q, q) is called the Lagrangian and is defined as L{q, q) = T(q, q) - W{ql i.e., the difference between the kinetic energy T and the potential energy W. The function D{q) denotes Rayleigh's dissipation function, which we shall assume to be of the form
^(^)- ^ i=\ la^iMp 7=1 where [jS/y] is a positive semidefinite matrix (i.e., [/3/y] is symmetric and all its eigenvalues are nonnegative). The dissipation function D represents one-half the rate at which energy is dissipated as heat. It is produced by friction in mechanical systems and by resistance in electric circuits. Finally, ft in Eq. (20.6) denotes an applied force and includes all external forces associated with the qt coordinate. The force ft is defined as being positive when it acts to increase the value of the coordinate qt. (a) In Fig. 1.21, M\ and M2 denote point masses; Ki, K2, K denote spring constants; yh yi denote the displacements of masses M\ and M2, respectively; and B\, B2, B denote viscous damping coefficients. Use the Lagrange formulation of dynamical
83 CHAPTER 1:
Mathematical Descriptions of Systems
84 Linear Systems
y77'Y/////yyyyy////y//7?^ 81
Bp
FIGURE 1.21 An example of a mechanical system with energy dissipation
systems described above to derive two second-order ordinary differential equations that characterize this system. Transform these equations into a system of first order ordinary differential equations. Verify your answer by using Newton's second law of motion to derive the same system equations. Assume that ji(0), ji(0), j2(0), J2(0) are given. (b) Consider the capacitor microphone depicted in Fig. 1.22. Here we have a capacitor constructed from a fixed plate and a moving plate with mass M. The moving plate is suspended from the fixed frame by a spring that has a spring constant k and also some damping expressed by the damping constant B. Sound waves exert an external force fit) on the moving plate. The output voltage v^, which appears across the resistor R, will reproduce electrically the sound-wave patterns that strike the moving plate. When fit) = 0 there is a charge q^ on the capacitor. This produces a force of attraction between the plates that stretches the spring by an amount x\, and the space between the plates is XQ. When sound waves exert a force on the moving plate, there will be a resulting motion displacement x that is measured from the equilibrium position. The distance between the plates will then be XQ - x, and the charge on the plates will be ^0 + ^•
Moving plate with mass M
-Xi -\- X-
K ->AAN— ^B
fit) -xo-x-
Fixed plate
tl'tr^
FIGURE 1.22 Capacitor microphone
•^Tjcir—'
V
When displacements are small, the expression for the capacitance is given approximately by XQ
— X
with Co = eA/xo, where e > 0 is the dielectric constant for air and A is the area of the plate. Use the Lagrange formulation of dynamical systems to derive two second-order ordinary differential equations that characterize this system. Transform these equations into a system of first-order ordinary differential equations. Verify your answer by using Newton's laws of motion and Kirchhoff's voltage/current laws. Assume that x(0), i(0), ^(0), and q{0) are given. (c) Use the Lagrange formulation to derive a system of first-order differential equations for the system given in Example 4.3. 1.3. Find examples of initial-value problems for which (a) no solutions exist; (b) more than one solution exists; (c) one or more solutions exist, but cannot be continued for all r G R\ and (d) unique solutions exist for all r G /?. 1.4. (Numerical solution of ordinary differential equations—Euler's method) In Subsection 1.6B it is shown that an approximation to the solution of the scalar initial-value problem fit, y), y = jyt,y),
j(ro) = = Jo JO y{to)
(20.7)
k =
(20.8)
is given by Euler's method [see Eq. (6.6)], yk+i = yk + hf{tk,yk\
Q,h2,.
where h = tk+\— tk is the (constant) integration step. The interpretation of this method is that the area below the solution curve [see Eq. (V) in Section 1.6] is approximated by a sequence of sums of rectangular areas. This method is also called \hQ forward rectangular rule (of integration), (a) Use Euler's method to determine the solution of the initial-value problem j = 3j,
y{tQ)
tQ = 0,
tQ^ t ^ 10.
(b) Use Euler's method to determine the solution of the initial-value problem y = t{yf -
/
yit^) = 1,
y(to) = 0,
to = 0,
to ^ t ^ 10.
Hint: In both cases, use h = 0.2. For part (b), let y = xi, xi = X2, X2 = tx^ - x\, and apply (20.8), appropriately adjusted to the vector case. In both cases, plot yk vs. ^^ y^ = 0,1, 2 , . . . . Remark: Euler's method yields arbitrarily close approximations to the solutions of (20.7), by making h sufficiently small, assuming infinite {computer) word length. In practice, however, where truncation errors (quantization) and round-off errors (finite precision operations) are a reality, extremely small values of h may lead to numerical instabilities. Therefore, Euler's method is of limited value as a means of solving initialvalue problems numerically. 1.5. (Numerical solution of ordinary differential equations—Runge-Kutta methods) The Runge-Kutta family of integration methods are among the most widely used techniques to solve initial-value problems (20.7). A simple version is given by yi+i = yi + K where
k = Uki -h 2y^2 + 2^3 + ^4),
85 CHAPTER 1 :
Mathematical Descriptions of Systems
86
with
Linear Systems
k\ = hf(ti, yt) fe
= hfiu + ^hyi
+ ^ki)
h = hf(ti + i/i,};/ + ife) k4 = hfiti + h, ji + h), andf/+i = ti + Kyito) = y^. The idea of this method is to probe ahead (in time) by one-half or by a whole step h to determine the values of the derivative at several points, and then to form a weighted average. Runge-Kutta methods can also be applied to higher order ordinary differential equations. For example, after a change of variables, suppose that a second-order differential equation has been changed to a system of two first-order differential equations, say, x\ = fi(t,xi,X2),
xiito) = xio
Xl = flit, Xi, X2X
X2(to) = X20.
^20 9)
In solving (20.9), a simple version of the Runge-Kutta method is given by yt+i = Ji + k, yi = {xu, xiiY and k = (k, lY
where with
k = ^('^i + ^h + 2fe + h),
I = l(l\+
2/2 + 2/3 + U)
and h = hfi{tuXii,X2i)
h = hfiiU, Xu, X2i)
h = hfiiti + i/z, Xu + ^ki, X2i + 5/1)
h = hf2(ti + ^h, xu + ^h, X2i + 5/1)
h = hfiiU + ^h, Xu + ^k2, X2i + 5/2)
h = hf2(ti + \K Xu + {h, X2i + ^2)
h = hfi(ti + h, Xu + ks, X2i + h)
k = hf2(ti + h, Xu + h, X2i + h\
Use the Runge-Kutta method described above to obtain numerical solutions to the initial value problems given in parts (a) and (b) of Exercise 1.4. As there, plot your data. Remark: Since Runge-Kutta methods do not use past information, they constitute attractive starting methods for more efficient numerical integration schemes (e.g., predictor-corrector methods). We note that since there are no built-in accuracy measures in the Runge-Kutta methods, significant computational efforts are frequently expended to achieve a desired accuracy. 1.6. (Numerical solution of ordinary differential equations—Predictor-Corrector methods) A common predictor-corrector technique for solving initial-value problems determined by ordinary differential equations, such as (20.7), is the Milne method, which we now summarize. In this method, yt-i denotes the value of the first derivative at time ^/_i, where ti is the time for the /th iteration step, j / - 2 is similarly defined, and yt+i represents the value of j to be determined. The details of the Milne method are: Ah, 1- yi+i,p = yi-3 + —(2J/-2 - yt-i + 2j/) (predictor)
T
2. yi+i,p =
fiti+uyt+ip) h 3. yt+ic = yt-i + 3 (i^z-i + 4j/ + yt+ip)
f(ti+uyi+i,c) h 5. yt+u = yi-i + 3 (>'/-i + 4j/ + yt+ic)
(corrector)
4. yi+i,c =
(iterating corrector)
The first step is to obtain a predicted value of yt+i and then substitute yi+\,p into the given differential equation to obtain a predicted value of yi+i, as indicated in the
second equation above. This predicted value, yi+i,p is then used in the second equation, the corrector equation, to obtain a corrected value of yi+\. The corrected value, yt+ic is next substituted into the differential equation to obtain an improved value of y/+i, and so on. If necessary, an iteration process involving the fourth and fifth equations continues until successive values of yi+i differ by less than the value of some desirable tolerance. With yt+i determined to the desired accuracy, the method steps forward one h increment. A more complicated predictor-corrector method that is more reliable than the Milne method is the Adams-Bashforth-Moulton method, the essential equations of which are yi+i,p
= yt +
Ji+U
^ 5 5 ^ , - 59y,-i + 31yi-2 - 9^,-3) 24'
J/ + - ( 9 j , + i + 19j,-
5 j / - i + yi-2),
where in the corrector equation, yi+\ denotes the predicted value. The application of predictor-corrector methods to systems of first-order ordinary differential equations is straightforward. For example, the application of the Milne method to the second-order system in (20.9) yields from the predictor step 4/z. ^k,i + \,p = Xk,i-3 + ^(2Xk,i-2
Then
^k,i+i,p
— fk{U+\,
- Xk,i^i
^i,i+\,py
+ 2Xk,il
k = 1, 2.
k=
X2,i+\,p)>
1,2,
and the corrector step assumes the form ^k,i+\,c
and
— ^k,i-l
+ l^{^k,i-\
+ ^^k,i
Xk,i+\,c = fk(ti+h xij+ic, X2,i+i,c\
+ ^k,i + \),
k — 1,2,
k = 1,2.
Use the Milne method and the Adams-Bashforth-Moulton method described above to obtain numerical solutions to the initial-value problems given in parts (a) and (b) of Exercise 1.4. To initiate the algorithm, refer to the Remark in Exercise 1.5. Remark. Derivations and convergence properties of numerical integration schemes, such as those discussed here and in Exercises 1.4 and 1.5, can be found in many of the standard texts on numerical analysis (see, e.g., D. Kincaid and W. Cheney, Numerical Analysis, Brooks-Cole, Belmont, CA, 1991). 1.7. (a) (b) (c) (d) (e) (f)
Prove Theorem Prove Theorem Prove Theorem Prove Theorem Prove Theorem Prove Theorem
6.2 for the interval [to - c, to]. 7.2 for the endpoint a. 8.2 for the interval [to - d, to]. 8.4 for t < to. 8.5 for the interval [to - c,to]. 10.1 through and including Theorem 10.10.
1.8. (a) Prove that the function f(x) = x^^^ is continuous but not Lipschitz continuous. (b) Show that the initial-value problem i; = x^^^, x(0) = 0 has infinitely many different solutions. (Remember to consider t < 0.) 1.9. Use Theorem 8.5 to solve the initial-value problem x = ax + t, x(0) = xo for ^ > 0. Here a E R. 1.10. Consider the initial-value problem X = Ax,
x(0) - Xo,
(20.10)
87 CHAPTER 1:
Mathematical Descriptions of Systems
88 Linear Systems
where x ^ R^ and A E R^^^. Let Ai, A2 denote the eigenvalues of A, i.e., Ai and A2 ^^^ ^^^ ^^^^^ ^^ ^^^ equation det (A - XI) = 0, where det denotes determinant, A is a scalar, and / denotes the 2 X 2 identity matrix. Make specific choices of A to obtain the following rollowing cases: cases: L 2. ^ 3. 4. 5. 6. 7. 8.
Ai Ai \. Ai Ai Ai Ai Ai Ai
> < == = > = = =
0, A2 > 0, and Ai 7^ A2; 0, A2 < 0, and Ai 7^ A2; \. > ^ 0; 0A2 A2 < 0; 0, A2 < 0; a + /j8, A2 = a - /jS, / = v ^ , a > 0; a + //3, A2 = a - //3, a < 0; /jS,A2 = - / / 3 .
Using r as a parameter, plot (l)2(t, 0, xo) vs. (t)i(t, 0, XQ) for 0 < ^ < r/ for every case enumerated above. Here [(/>i(t, to, XQ), (f)2{t, to, xo)V = rrra/r5 (see Exercise LIO) for this example for the cases e = 0.05 and e = 10 (refer also to Exercises L5 and L6 for numerical methods for solving differential equations). The periodic function to which the trajectories of (4.10) tend is an example of a limit cycle. 1.12. For (4.11) consider the hard, linear, and soft spring models given by g(x) = k(l + a^x^)x, g{x) = kx, g(x) = k(l - a^x^)x, respectively, where ^ > 0 and a 7^ 0. Write two first-order ordinary differential equations for (4.11) by choosing x\ = x and X2 = x. Pick specific values for k and a^. Determine by simulation/7/i(2>y^ portraits (see Exercise 1.10) for this example for the above three cases. 1.13. Verify that the spaces in Examples 10.1 to 10.6 satisfy the axioms of vector space. 1.14. Let F = {0, 1, 2, 3}. Determine operations " + " and "•" so that [F, -H, •} is a field. 1.15. (a) Verify that the transformation given in Example 10.8 is linear, (b) Verify that the transformation given in Example 10.9 is linear. 1.16. (a) Prove (10.13), (10.14), (10.15), and (10.16). (b) Verify that the functions || • ||i, || • lb, and || • |oo defined in (10.11), (10.12), and (10.10), respectively, each satisfy the axioms of a norm. (c) Prove the relations (D-i), (D-ii), (D-iii) given at the beginning of Subsection 1 .IOC. 1.17. Let g{t) = [g\(t),..., gn(t)V be defined on some interval J C R and assume that gi : J ^ R is integrable over J,i = I,.. .,n. Prove that for b > a, \\\^ g(t)dt\\ < J 11^(011 dt, where || • || denotes a norm on 7?".
1.18. (a) Show that x^ = (0, 0) is a solution of the system of equations CHAPTER 1:
Xi = x\ + x\ + X2 cos X\
X2 = {\ + X\)x\ + (1 + X2)x2 + xi sinx2. Linearize this system about the point x^ = (0, 0). By means of computer simulations, compare solutions corresponding to different initial conditions in the vicinity of the origin of the above system of equations and its linearization. (b) Linearize the (bilinear control) system jc + (3 + x^)x + (1 + jc + x^)u = 0 about the solution ;c = 0, i = 0, and the input u{t) = 0. As in part (a), compare (by means of computer simulations) solutions of the above equation with corresponding solutions of its linearization. (c) In the circuit given in Fig. 1.23, v/(0 is a voltage source and the nonlinear resistor obeys the relation IR = 1.5v| [Vi(t) is the circuit input and v/?(0 is the circuit output]. Derive the differential equation for this circuit. Linearize this differential equation for the case when the circuit operates about the point v/ = 14. 'fl(0
1Q
-MAr
.(.) Q
1 F.
VR(t)
FIGURE 1.23 Nonlinear circuit 1.19. Consider a system whose state-space description is given by X = —k\k2j^+
k2u(t)
y = ki Jx. Linearize this system about the nominal solution Wo = 0,
2 Vxo(0 = 2 Jk — ki k2t,
where xo(0) = k. 1.20. (Inverted pendulum) The inverted pendulum on a moving carriage subjected to an external force ix{t) is depicted in Fig. 1.24. The moment of inertia with respect to the center of gravity is / and the coefficient of friction of the carriage (see Fig. 1.24) is F. From Fig. 1.25 we obtain the following equations for the dynamics of this system: m—r(5' + Lsin(/>) = H
(20.11a)
Mathematical Descriptions of Systems
90 Linear Systems
(Center of gravity)
FIGURE 1.24 Inverted pendulum
m-r^{Lcos^) dfi
= Y - mg
(20.11b)
J—p2 ^ LY sin (j) - LH COS (I)
(20.11c) (20. lid)
Assuming that m «
M, (20.lid) reduces to ,^d^S
,,
dS
(20. lie)
Eliminating H and Y from (20.11a) to (20.11c), we obtain (/ + mL^)cf> = mgL sin (/) - mLS cos (/>.
(20.1If)
Thus, the system of Fig. 1.24 is described by the equations 4> - (77 )sin(^ + (^]scos(t>
= 0,
MS + F 5 = fji(t),
FIGURE 1.25
(20.11g)
L' =
where
denotes the effective pendulum length. Linearize system (20.11g) about cj)
91
J + mL'^ mL
CHAPTER 1:
Mathematical Descriptions of Systems
0.
1.21. (Magnetic ball suspension system) Figure 1.26 depicts a schematic diagram of a ball suspension control system. The steel ball is suspended in air by the electromagnetic force generated by the electromagnet. The objective of the control is to keep the steel ball suspended at a desired equilibrium position by controlling the current i(t) in the magnet coil by means of the applied voltage v(0, where r > 0 denotes time. The resistance and inductance of the coil are R and L(s(t)) = L/s(t), respectively, where L > 0 is a constant and s(t) denotes the distance between the center of the ball and the magnet at time t. The force produced by the magnet is Kfl(t)/s^(t), where jfiT > 0 is a proportional constant and g denotes acceleration due to gravity. (a) Determine the differential equations governing the dynamics of this system. (b) Let v(t) = Veq,2i nominal (desired) value of v{t). Determine the resulting equilibrium of this system. (c) Linearize the equation obtained in (a) about the equilibrium solution determined in (b).
y//////////^ NsN\^ R
ns^S^—^*--o"^ L v{t)
Electromagnet
Steel ball
FIGURE 1.26 Magnetic ball suspension system 1.22. (a) For the mechanical system given in Exercise 1.2a, we view / i and / i as making up the system input vector, and y\ and y2 the system output vector. Determine a state-space description for this system. (b) For the same mechanical system, we view (/i, 5/2)^ as the system input and we view 83^1 + 10^2 as the (scalar-valued) system output. Determine a state-space description for this system. (c) For part (a), determine the input-output description of the system. (d) For part (b), determine the input-output description of the system.
92 Linear Systems
1.23. For the magnetic ball suspension system given in Exercise 1.21, we view v and s as the system input and output, respectively. (a) Determine a state-space representation for this system. (b) Using the linearized equation obtained in part (c) of Exercise 1.21, obtain the inputoutput description of this system. 1.24. In Example 4.3, we view Ca and 9 as the system input and output, respectively. (a) Determine a state-space representation for this system. (b) Determine the input-output description of this system. 1.25. For the second-order section digital filter in direct form, given in Fig. 1.13, determine the input-output description, where xi(k) and u(k) denote the output and input, respectively. 1.26. In the circuit of Fig. 1.27, Vi{t) and Vo(t) are voltages (at time t) and Ri and R2 are resistors. There is also an ideal diode that acts as a short circuit when V/ is positive and as an open circuit when Vi is negative. We view Vi and VQ as the system input and output, respectively. (a) Determine an input-output description of this system. (b) Is this system Hnear? Is it time-varying or time-invariant? Is it causal? Explain your answers.
Diode
^•(0
i{t)
O
^o(0
FIGURE 1.27 Diode circuit 1.27. We consider the truncation operator given by y{t) = Tr(u(t)) as a system, where r E 7? is fixed, u and y denote system input and output, respectively, t denotes time, and Tr( • ) is specified by Tr(u(t))
u{t), 0,
t < T, t > T.
Is this system causal? Is it linear? Is it time-invariant? What is its impulse response? 1.28. We consider the shift operator given by y{t) = Qr(u(t)) = u(t - T) as a system, where r ^ R is fixed, u and y denote system input and system output, respectively, and t den tes time. Is this system causal? Is it linear? Is it time-invariant? What is its impulse response?
1.29. Consider the system whose input-output description is given by
CHAPTER 1:
y(t) = min{wi (0,^2(01, where u(t) = [ui(t), U2(t)]^ denotes the system input and y(t) is the system output. Is this system linear? 1.30. Suppose it is known that a linear system has impulse response given by h(t, r) = exp(-|r - T|). IS this system causal? Is it time-invariant? 1.31. Consider a system with input-output description given by y(k) = 3u(k + 1) + 1,
kGZ,
where y and u denote the output and input, respectively (recall that Z denotes the integers). Is this system causal? Is it linear? 1.32. Use expression (16.8), x(n) = ^
x(k)8(n - k),
and 8(n) = p(n) - p(n - 1) to express the system response y(n) due to any input u{k), as a function of the unit step response of the system [i.e., due to u{k) = p(k)]. 1.33. (Simple pendulum) A system of first-order ordinary differential equations that characterize the simple pendulum considered in Exercise 1.1b is given by X2 — J smjJCi
where xi = 6 and X2 = 6 with xi(0) = 0(0) and ^2(0) = 6(0) specified. A linearized model of this system about the solution x = [0, 0]"^ is given by Xi X2\
0
11
1 °
93
Xi X2\
Let ^ = 10 (m/sec^) and / = 1 (m). (a) For the case when x(0) = [OQ, Of with ^o = ^/18, 77/12, 7r/6, and 7r/3, plot the states for t > 0, for the nonlinear model. (b) Repeat (a) for the linear model. (c) Compare the results in (a) and (b).
Mathematical Descriptions of Systems
CHAPTER 2
Response of Linear Systems
In system theory it is important to clearly understand how inputs and initial conditions affect the response of a system. There are many reasons for this. For example, in control theory, it is important to be able to select an input that will cause the system output to satisfy certain properties [e.g., to remain bounded (stability), to follow a given trajectory (tracking), and the like]. This is in stark contrast to the study of ordinary differential equations, where it is usually assumed that the forcing function (input) is given. 2.1 INTRODUCTION The goal of this chapter is to study the response of linear systems in greater detail than was done in Chapter 1. To this end, solutions of linear ordinary differential equations are reexamined, this time with an emphasis on characterizing all solutions using bases (of the solution vector space) and on determining such solutions. For convenience, certain results from Chapter 1 are repeated, and time-varying and time-invariant cases are treated separately, as are continuous-time and discrete-time cases. Whereas in Chapter 1 certain fundamental issues that include input-output system descriptions, causality, linearity, and time-invariance are emphasized, we will here address in greater detail impulse (and pulse) response and transfer functions for continuous-time systems and discrete-time systems. A. Chapter Description As in Chapter 1, we will concern ourselves with linear, continuous-time, finitedimensional systems represented by the state and output equations (internal description) 94
X = A(t)x + B(t)u,
y = C(t)x + D(t)u
(1.1)
CHAPTER 2 :
and the corresponding input-output description (external description) H{t,T)u{T)dT,
y(t)
(1.2)
where H{t, r) denotes the impulse response matrix. As pointed out in Chapter 1, in the special case when A{t) = A, B(t) = B, C{t) = C, and D(t) = D, with initial time to = 0, the external description (1.2) can be represented equivalently in terms of Laplace transform variables y(s) = H(s)u(s),
(1.3)
where H(s) denotes the transfer function matrix of the system. We will examine in greater detail the relationship between the internal description (1.1) and the external description (1.2) here and in subsequent chapters. Again, as in Chapter 1, we will also concern ourselves with linear, discrete-time, finite-dimensional systems represented by the state and output equations (internal description) x(k + 1) = A(k)x(k) + B(kMk),
y(k) - C(k)x(k) + D(k)u(k)
(1.4)
and the corresponding input-output description (external description) n
y(n) = ^H{n,k)u{k\
(1.5)
k= h
where H{n, k) denotes the unit pulse response. In the time-invariant case, with k^ = 0, (1.5) can be represented equivalently (in terms of z-transform variables) as y(z) = H(z)u(zl
95
(1.6)
where H(z) denotes the system transfer function matrix. In the following, we provide a brief outline of the contents of this chapter. In the second section we provide some mathematical background material on linear algebra and matrix theory. In the third section we further study systems of linear homogeneous and nonhomogeneous ordinary differential equations. Specifically, in this section we develop a general characterization of the solutions of such equations and we study the properties of the solutions by investigating the properties of fundamental matrices and state transition matrices. In section four we further investigate systems of linear, autonomous, homogeneous ordinary differential equations. In particular, in this section we emphasize several methods of determining the state transition matrix of such systems and we study the asymptotic behavior of the solutions of such systems. In the fifth section we study systems of linear, periodic ordinary differential equations (Floquet theory). In the sixth and seventh sections we further investigate the properties of the state representations and the input-output representations of continuous-time and discrete-time finite-dimensional systems. Specifically, in these sections we study equivalent representations of such systems, we investigate the properties of transfer function matrices, and for the discrete-time case we also address sampleddata systems and the asymptotic behavior of the system response of time-invariant systems.
Response of Linear Systems
96 Linear Systems
B. Guidelines for the Reader In a first reading, the background material on linear algebra, given in Section 2.2, can be reviewed rather quickly. In the study of the subsequent material of this book, selective detailed coverage of topics in Section 2.2 may also be desirable. A typical beginning graduate course in linear systems will include Theorem 3.1 in Subsection 2.3 A, which shows that the set of solutions of the linear homogeneous equation x = A(t)x forms an n-dimensional vector space. This theorem provides the basis for the definitions of the fundamental and the state transition matrix (Subsections 2.3A, 2.3B, and 2.3D). These results enable us to determine solutions of the nonhomogeneous equation x = A(t)x + g(t) in Subsection 2.3C. Background required for the above topics includes material on vector spaces, linear independence of vectors, bases for vector spaces, and linear transformations (Subsections 2.2A to 2.2E). A typical beginning graduate course in linear systems will also address the material on time-invariant systems given in Section 2.4, including solutions of the equation X = Ax-\- g(t), various methods of determining the matrix exponential e^\ and asymptotic behavior of time-invariant systems x = Ax (including system modes and stability properties of an equilibrium). Background required for these topics includes material on equivalence and similarity, eigenvalues and eigenvectors (Subsections 2.21 and 2.2J). In addition to the above, a beginning graduate course on linear systems will also treat the input-output description of continuous-time systems and its relation to state-space representations, including the impulse response for time-varying and time-invariant systems, the transfer function matrix for time-invariant systems, and the equivalence of state-space representations (Section 2.6). Finally, such a course will also cover state-space and input-output descriptions of discrete-time systems (Section 2.7).
2.2 BACKGROUND MATERIAL In this section we consider material from linear algebra and matrix theory. We assume that the reader has some background in these areas, and therefore, our presentation will constitute a summary rather than a development of the subject matter on hand. This section consists of fifteen subsections. In the first three subsections we consider Hnear subspaces of vector spaces, linear independence of a set of vectors, and bases of vector spaces, respectively. In the next five subsections, we address general linear transformations defined on vector spaces, the representation of such transformations by matrices, some of the properties of matrices and determinants of matrices, and solutions of linear algebraic equations, respectively. In the ninth and tenth subsections, we address equivalence and similarity of matrices and eigenvalues and eigenvectors, respectively. In the eleventh subsection we digress by considering direct sums of linear subspaces. In the last four subsections we address, respectively, certain canonical forms of matrices, minimal polynomials of matrices, nilpotent operators, and the Jordan canonical form.
A. Linear Subspaces
97
In Section 1.10 we gave the formal definition of vector space over afield, say, (K F), where V denotes the set of vectors and F denotes the set of scalars. When F (the field) is clear from context, we usually speak of a vector space V (or a linear space V) rather than {V, F). A nonempty subset W of a vector space V is called a linear subspace (or a linear manifold) in V if (i) w\ + W2 is in W whenever w\ and W2 are in W, and (ii) aw is in W whenever a G F and w ^ W.li is an easy matter to verify that a linear subspace W satisfies all the axioms of a vector space and may as such be regarded as a linear space itself. Two trivial examples of linear subspaces include the null vector (i.e., the set W = {0} is a linear subspace of V) and the vector space V itself. Another example of a linear subspace is the set of all real-valued polynomials defined on the interval [a, b], that is a linear subspace of the vector space consisting of all real-valued continuous functions defined on the interval [a, b] (refer to Example 10.4 in Chapter 1). As another example of a linear subspace (of R^), we cite the set of all points on a straight line passing through the origin. On the other hand, a straight line that does not pass through the origin is not a linear subspace of R^ (why?). It is an easy matter to show that if Wi and W2 are linear subspaces of a vector space V, then Vl^i Pi 1^2. the intersection of Wi and W2, is also a linear subspace of V. A similar statement cannot be made, however, for the union of Wi and W2 (prove this). Note that to show that a set V is a vector space, it suffices to show that it is a linear subspace of some vector space.
Response of Linear Systems
CHAPTER 2 :
B. Linear Independence Throughout the remainder of this section, we let { a i , . . . , a„}, at G F, denote an indexed set of scalars and we let {v\ . . . , v"}, v' E V, denote an indexed set of vectors. Now let W be a set in a linear space V (W may be a finite set or an infinite set). We say that a vector v E V is 3. finite linear combination of vectors in W if there is a finite set of elements {w^,..., w^] in W and a finite set of scalars { a i , . . . , an) in F such that V = a\w^ + •• • + a^w^. Now let Ty be a nonempty subset of a linear space V and let S{W) be the set of all finite linear combinations of the vectors from W, i.e., w G S{W) if and only if there is some set of scalars {a\,..., a^} and some finite subset {w^,..., w^} of W such that w = aiw^ ^ h a^v^^, where m may be any positive integer. Then it is easily shown that S(W) is a linear subspace of V, called the linear subspace generated by the set W. Now if t/ is a linear subspace of a vector space V and if there exists a set of vectors W CV such that the linear space S{W) generated by W is U, then we say that W spans U. It is easily shown that S(W) is the smallest Hnear subspace of a vector space V containing the subset W of V. Specifically, if t/ is a linear subspace of V and if U contains W, then U also contains S(W).
98 Linear Systems
As an example, in the space (R^, R) the set Si = {e^} = {(1, 0)^} spans the set consisting of all vectors of the form (a, 0)^, a EL R, while the set ^2 = {e^, e^}, e^ = (0,1)^ spans all of 7^2. We are now in a position to introduce the notion of linear dependence. DEFINITION 2.1. Let 5 = {v\ . . . , v'"} be a finite nonempty set in a linear space V. If there exist scalars ai,..., a^, not all zero, such that aiv^ + ••• + a ^ v ^ = 0,
(2.1)
then the set S is said to be linearly dependent (over F)- If a set is not linearly dependent, then it is said to be linearly independent. In this case relation (2.1) implies that a\ = • • • = am = 0. An infinite set of vectors T^ in V is said to be linearly independent if every finite subset of W is linearly independent. • EXAMPLE 2.1. Consider the linear space (R^,R) (see Example 10.1 in Chapter 1), and let e^ = (1, 0 , . . . , 0)^, e^ = (0, 1, 0 , . . . , 0 ) ^ , . . . , ^" = ( 0 , . . . , 0,1)^. Clearly, S"^i <^/^' = 0 implies that at = 0, / = I,.. .,n. Therefore, the set 5" = {e^,..., ^"} is a linearly independent set of vectors in R^ over the field of real numbers R. • EXAMPLE 2.2. Let V be the set of 2-tuples whose entries are complex-valued rational functions over the field of complex-valued rational functions. Let s+2 1 (s + 1)(^ + 3) s-h 1 1 1 ^s + 2 s +3 and let « ! == - 1 , 0^2 = (s + 3)/(s + 2). Then aiv^ + a^v^ = 0, and therefore, the set S = {v\ v^} is linearly dependent over the field of rational functions. On the other hand, since aiv^ +0:2 v^ = 0 when a i , 0:2 ^ ^ is true if and only if a i = 0:2 = 0, it follows that S is linearly independent over the field of real numbers (which is a subset of the field of rational functions). This shows that linear dependence of a set of vectors in V depends on the field F. m Linear independence of functions of time EXAMPLE 2.3. Let V = C{{a, b), R""), let F = R, and for x,y EV mda E F, define addition of elements in V and multiplication of elements in V by elements in F by (x + y)(t) = x(t) + y(t) for all t G (a, b) and iax){t) = ax(t) for all t E (a, b). Then as in Example 10.4 of Chapter 1, we can easily show that (V, F) is a vector space. An interesting question that arises is whether for this space, linear dependence (and linear independence) of a set of vectors can be phrased in some testable form. The answer is yes. Indeed, it can readily be verified that for the present vector space (K F), linear dependence of a set of vectors S = {(f)i,..., (f)k} in V = C{{a, b\ 7?") over F = Ris equivalent to the requirement that there exist scalars a/ G F, / = 1 , . . . , ^, not all zero, such that ai4>i(t) + • • • + ak(i)k(t) = 0
for all t G (a, b).
Otherwise, S is linearly independent. To see how the above example applies to specific cases, let V = C((-^, 00), R^) and consider the vectors (/)i(0 = [1, tY, b (f>2} is linearly independent (over F = R), assume for purposes of contradiction that S is linearly dependent. Then there must exist scalars ai and a2, not both zero, such that Q!i[l, tf + a2[l, t^f = [0, 0]^ for all t G ( - ^ , 00). But in particular, for t = 2, the above equation is satisfied if and only if a 1 = a2 = 0, which contradicts the assumption. Therefore, 5" = {(^1, (/)2} is linearly independent.
As another specific case of the above example, let V = C((-<^, ^), R^) and consider 99 the set S = {>!, ^2, h> >4}, where cj^iit) = [1, tf, (j)2{t) = [1, fl (jy^it) = [0, 1]^, andCHAPTER 2 : 04(0 = [e~^, 0]. The set S is clearly independent over R since aiipiit) + 0L2(t>2{t) +Response of 0^303(0 + 0L4r4>A{t) = 0 for all t G (-00, 00) if and only if ai = 0:2 = 0:3 = 0:4 = 0. •Linear Systems Next, let 5 = {v^ . . . , v^} be a linearly independent set in a vector space V. If Xr= 1 ^^^^ ^ 2 r = 1 i^i^^ t^^^ it is readily shown that at = j8/, for all / = 1 , . . . , m. Also, it is easily shown that the set S is linearly dependent if and only if for some index /, 1 < / < m, we can find scalars y i , . . . , 7/-1, 7/+1,..., ym such that v' = yiv^ H hy^~^v'~^ +y^+iyi+i -^ hy^v'^. Furthermore, it is not hard to verify that a finite nonempty set V^ in a linear space is linearly independent if and only if for each v G 5(W), v 7^ 0, there is a unique finite subset of W, say, {v^ v^,..., v^} and a unique set of nonzero scalars {81,..., S^}, such that v = Si v^ + • • • + 8mV^. Finally, if t/ is a finite set in a linear space V, then it is easily shown that U is linearly independent if and only if there is no proper subset Z of U such that S(U) = S(Z). (Recall that Z is a proper subset of U if there is a w G 6^ such that u ^Z.)
C. Bases We are now in a position to introduce another important concept. DEFINITION 2.2. A set W in a linear space V is called a basis for V if (i) W is linearly independent, (ii) The span of W is the linear space V itself, i.e., S(W) = V.
•
An immediate consequence of the above definition is that if W is a linearly independent set in a vector space V, then T^ is a basis for S(W). To introduce the notion of dimension of a vector space, it is shown that if a linear space V is generated by a finite number of linearly independent elements, then this number of elements must be unique. The following results lead up to this. Let {v^ . . . , v"} be a basis for a linear space V. Then it is easily shown that for each vector v EV, there exist unique scalars a\,. ..,an such that
Furthermore, ifu^,...,u^is any linearly independent set of vectors in V, then m < n. Moreover, any other basis of V consists of exactly n elements. These facts allow the definitions given in the following. If a linear space V has a basis consisting of a finite number of vectors, say, {v^ . . . , v'^}, then V is said to be di finite-dimensional vector space and the dimension of V is n, abbreviated dim V = n.ln this case we speak of an n-dimensional vector space. If V is not a finite-dimensional vector space, it is said to be an infinitedimensional vector space. By convention, the linear space consisting of the null vector is finite-dimensional with dimension equal to zero. An alternative to the above definition of dimension of a (finite-dimensional) vector space is given by the following result, which is easily verified: let V be a vector space that contains n linearly independent vectors. If every set of n + 1 vectors in V is linearly dependent, then V is finite-dimensional and dim V = n.
100 Linear Systems
The preceding results enable us now to introduce the concept of coordinates of a vector. We let {v^ . . . , v'^} be a basis of a vector space V and let v E V be represented by
The unique scalars ^ i , . . . , ^^^ are called the coordinates ofv with respect to the basis EXAMPLE 2.4. For the linear space (/?^/?), let S = {e\...,e''}, where the e' G R^, i = 1,..., n, were defined earlier (following Defintion 2.1). Then S is clearly a basis for (/?", R) since it is linearly independent and since given any v G R^, there exist unique real scalars at, i = 1,..., n, such that v = X"=i ocie^ = (ai,..., a:„)^, i.e., S spans /?". It follows that with every vector v E /?^, we can associate a unique n-tuple of scalars ai
or
(ai, ...,a„)
relative to the basis {e^,..., e^}, the coordinate representation of the vector v E /?" with respect to the basis 5" = {e\ ..., ^"}. Henceforth, we will refer to the basis S of this example as the natural basis for R^. • EXAMPLE2.5. We note that the vector space of all (complex-valued) polynomials with real coefficients of degree less than n is an ^-dimensional vector space over the field of real numbers. A basis for this space is given by S = {I, s,..., s'^~^}, where ^ is a complex variable. Associated with a given element of this vector space, say, p(s) = ao + ais + ' • • + an-\s^~^, we have the unique fz-tuple given by (ao, OL\, ..., ocn-iY, which constitutes the coordinate representation of p{s) with respect to the basis S given above. • EXAMPLE 2.6. We note that the space (V, /?), where V = C([a, b], R), given in Example 10.4 of Chapter 1, is an infinite-dimensional vector space (why?). •
D. Linear Transformations In Subsection 1.IDA we introduced the notion of linear transformation ST from a vector space V (over the fiield F) into a vector space W (over the same fiield F). Henceforth, we will write ST E L(V, W) to express this. It is our objective in this subsection to identify some of the important properties of linear transformations. Linear equations With Sr E L(V; ^ ) we define the null space of 3' as the set J<(^) = {v E y : STv = 0} and the range space of?) as the set 91(9") = {w E W : w = STv, V E y}. Note that since STO = 0, }({^) and 2/1(9") are never empty. It is easily verified that M{^) is a linear subspace of V and that 91(9") is a linear subspace of W. If V is finite-dimensional (of dimension n), then it is easily shown that dim 91(9") < n. Also, if V is finite-dimensional and if {w^ . . . , w^} is a basis for 91(9") and v^ is defined
by ^V w i = I,.. .,n, then it is readily proved that the vectors v^ . . . , v^ are Hnearly independent. One of the important results of linear algebra, called iht fundamental theorem of linear equations, states that for ^ G L{V, W) with V finite-dimensional, we have dim>r(2r) + dimSlCST) = dim V. For the proof of this result, refer to any of the references on linear algebra cited at the end of this chapter. The above result gives rise to the notions of the rank, p{^), of a linear transformation ST of a finite-dimensional vector space V into a vector space W, which we define as the dimension of the range space 9l(2r), and the nullity, v{?F), of ST, which we define as the dimension of the null space }^{^). With the above machinery in place, it is now easy to establish the following important results concerning linear equations. We let ST E L{V, W), where V is finitedimensional, let s = dim>r(2r), and let {v^ . . . , v^} be a basis for >r(2r). Then it is easily verified that (i) a vector v G V satisfies the equation STv = 0 if and only if V = XJ= 1 <^/v^ for some set of scalars {ai,..., as}, and furthermore, for each v G V such that STv = 0 is true, the set of scalars {ai,..., a^} is unique; (ii) ifw^GW is a fixed vector, then STv = w^ holds for at least one vector v G V (called the solution of the equation STv = w^) if and only if w^ G 5/1(9"); and (iii) if w^ is any fixed vector in W and if v^ is some vector in V such that STv^ = w^ (i.e., v^ is a solution of the equation STv^ = w^), then a vector v G V satisfies STv = w^ if and only if V = v^ + X / = i PiV^ for some set of scalars {/3i,..., Ps}, and furthermore, for each V G V such that STv = WQ, the set of scalars {j8i,..., jS^} is unique. General properties of linear transformations Before proceeding further, we briefly digress by recalling certain elementary properties of a function/ from a set X to a set 7, written f : X ^ Y. Letting 2/l(/) denote the range of/, we can classify/ in the following manner: if 9l(/) = F, then / is said to be surjective or a surjection and we say t h a t / maps X onto Y; iff is such that for every xi, X2 G X, f(xi) = f(x2) implies that xi = X2, t h e n / is said to be injective or an injection or a one-to-one mapping; and if/ is both injective and surjective, we say that/ is bijective or a one-to-one and onto mapping or a bijection. W h e n / is injective, its inverse f'^ : 2/l(/) -^ X exists, so that f~^(f(x)) = x for all X G X and f{f~^(y)) = y for all y G 9l(/). Note that w h e n / is bijective, we h a v e / - I :Y^ X. Returning now to the subject on hand, we note that since a linear transformation ST of a linear space V into a linear space W is a mapping, we distinguish in particular among linear transformations that are surjective, injective, and bijective. We will often be particularly interested in knowing when a linear transformation ST has an inverse 2r~^ When this is the case, we say that 9" is invertible, that ST"^ exists, or that 9" is nonsingular. A linear transformation that is not nonsingular is said to be singular. Concerning the inverse of a linear transformation ST G L(V, W), it is easily shown that ST"^ exists if and only if STv = 0 implies v = 0, and furthermore, if 2r~^ exists, then ST"^ is a linear transformation from 9l(2r) onto V [i.e., S''^ G L(2ft(2r), V)]. Moreover, if V is finite-dimensional, then S" has an inverse if and only if 91(9") has the same dimension as V, i.e., p(9') = dim V. Also, if both
101 CHAPTER 2 :
Response of Linear Systems
102 Linear Systems
V and W are finite-dimensional and of the same dimension, then 9l(2r) = W if and only if ST has an inverse. In the next few results, which are phrased in terms of equivalent statements, we summarize some of the important properties of linear transformations. (Injective linear transformations) For 9" G L{V, W), the following statements are equivalent: (i) ST is injective; (ii) ST has an inverse; (iii) STv = 0 implies V = 0; (iv) for each w G ?k('3'), there is a unique v G V such that STv = w; (v) if gTv^ = STv^ then v^ - v^; (vi) if v^ T^ V^ then STv^ T^ STV^. If in addition, V is finite-dimensional, then the following are equivalent: (i) ST is injective; (ii) p(2r) = dim V. (Surjective linear transformations) For ST G L{V, W), the following statements are equivalent: (i) ST is surjective; (ii) for each w G W, there is a v G V such that STv = w. If in addition, V and W are finite-dimensional, then the following are equivalent: (i) ST is surjective; (ii) dim W = pi^f). (Bijective linear transformations)¥or3' G L(V, W), the following are equivalent: (i) 9" is bijective; (ii) for every w G W, there is a unique v E V such that 9'v = w. If in addition y and Ware finite-dimensional, then the following are equivalent: (i) ST is bijective; (ii) dim V = dim W = p(?J'). (Injective, surjective, and bijective linear transformations) For S' E L(V, W) with V and W finite-dimensional and with dim V = dim W, the following are equivalent: (i) ST is injective; (ii) ST is surjective; (iii) ST is bijective; (iv) ST has an inverse. Next, we examine some of the properties of L(V, W), the set of all linear transformations from a vector space V into a vector space W. As before, we assume that V and W are linear spaces over the same field F. We let ^ , 3" G L(V, W) and we define the sum of^ and ST by (^ + 2r)v = SPv + STv
(2.2)
for all V G y. Also, with a G F and ST G L(V; W), we define multiplication of ST by a scalar a as {a^)v
= a^v
(2.3)
for all V G y. It is easily shown that (&'-\-^)G L(V, W) and also that aST G L(V, W). We further note that there exists a zero element in L(V, W), called the zero transformation, denoted by 0 and defined by Cv = 0
(2.4)
for all V G y. Furthermore, we note that to each ST G L(V, W) there corresponds a unique linear transformation - 9 " G L(V, W) defined by (-oj-)v ^ -STv
(2.5)
for all V G y. In this case it follows trivially that -ST + ST = 0. With these definitions in place, it is easily proved that L(V, W) is a linear space over F, called the space of linear transformations [with vector addition defined by (2.2) and multiplication of vectors by scalars defined by (2.3)]. To explore the properties of the space of linear transformations further, we briefly digress to recall the definition of an algebra. Specifically, a set V is called an algebra
if it is a linear space and if in addition to each v,w GV there corresponds an element in V, denoted by v • w and called the product ofv times w, satisfying the following axioms: 1. V • (w + w) = V • w + V • w for all V, w, w G y. 2. (v + w) • u = V ' u -\- w ' u for dill V, w, u E: V. 3. (av)' (f3w) = (cj^/3)(v • w) for all v, w E V and for all a, p G F. If in addition to the above, 4. (v'w)'
u = V'(w- u) for all v,w,uG
V, then V is called an associative algebra.
If there exists an element / G V such that i-v = v i = v for every v G V, then / is called the identity of the algebra. It can readily be shown that if / exists, then it is unique. Furthermore, if v • w = w • v for all v, w G V, then V is said to be a commutative algebra. Returning to the subject on hand, let V, W, and U be linear spaces over F, and consider the vector spaces L(V, W) and L(W, U). If S/^ G L{W, U) and ST G LiV, W), then we define the product 5f ST as the mapping of V into U by the relation (^2r)v = ^(STv)
(2.6)
for all V G y. It is easily verified that ^ST G L{V, U). Next, let y = \y = f/. If ^,^,^E LiV, V) and if a, jS G F, then it is easily shown that
^(sra) = (SP2r)a
and
(^ + 2r)a = jf a + gra (aSf)(/32r) = (af3)&'^.
(2.7) (2.8) (2.9) (2.10)
We emphasize that, in general, commutativity of linear transformations does not hold, i.e., in general
spsr 7^ srsp.
(2.11)
There is a special mapping from a linear space V into V, called the identity transformation, defined by 3v = V
(2.12)
for all V G y. We note that J^ G L(V, V), that ^ T^ 0 if and only if V T^ {0}, that 3 is unique, and that aj-^ = ^Gj- ^ aj(2.13) for all ST G L(V, V). Also, we can readily verify that the transformation a J^, a G F, defined by (a^)v = aS>v = av
(2.14)
is also a linear transformation. Relations (2.7) to (2.14) now give rise to the following result: L(V,V) is an associative algebra with identity ^. This algebra is in general not commutative.
103 CHAPTER 2 :
Response of Linear Systems
104 Linear Systems
Concerning invertible linear transformations we note that if J G L(V; V) is bijective, then ST"^ G L(V, V), and furthermore, of-^oj- = g-gj-1 ^ ^^ (2.15) where ^ denotes the identity transformation defined in (2.12). Next, if V is a finite-dimensional vector space and ST G L(V, V), then we can readily show that the following are equivalent: (i) ST is invertible; (ii) pCJ) = dim V; (iii) ST is one-to-one; (iv) ST is onto; and (v) STv = 0 implies that v = 0. For bijective linear transformations, we can easily verify the following characterizations. Let £/", ST, a G L(V, V) and let ^ denote the identity transformation. Then (i) if SPST = aSP - J^, then £P is bijective and Sf ~i = ST = 1; (ii) if ^ and Sr are bijective, then if?r is bijective, and (S^^y^ = ^T'^if-^; (iii) if ^ is bijective, then (S^'^y^ = ^ ; and (iv) if ^ is bijective, then a ^ is bijective and (aSP)-i = ( l / a ) y - i for all a G F, a T^ 0. With the aid of the above concepts and results we can now construct certain classes of functions of linear transformations. Since (2.7) allows us to write the product of three or more linear transformations without the use of parentheses, we can define 3'^, where 9" G L(V, V) and n is a positive integer, as aj-n A Of ,0}- . .,.
.oj-^^
(2.16)
n times
Similarly, if ST Ms the inverse of ST, then we can define 9" ^, where m is a positive integer, as
aj-m A (2r-i)^ =. ^-^ ^or-\' '" -gr-y
(2.17)
m times
Using these definitions, the usual laws of exponents can be verified. Thus, aj-m ,aj-n ^ /(jrm\n
and
oj-m+n ^
aj-n , aj-m
__ oj-mn __ arnm _ rarnyn
ST"' • 9"-" = gT'"-",
(2.I8) /o 1 Q \
(2.20)
where m and n are positive integers. Consistent with the above, we also have and
3"! = Sr
(2.21)
Sro = 3>.
(2.22)
We are now in a position to consider polynomials of linear transformations. For example, if /(A) is a polynomial, i.e., /(A) = ao + aiA + ---+a„A", where oio,..., Q!^ G i^, then by f(^)
(2.23)
we mean
/(ST) = a o i + aiSr + • • • + A^gr^
(2.24)
The reader is cautioned that in general the above concept cannot be extended to functions of two or more linear transformations, because linear transformations in general do not commute. E. Representation of Linear Transformations by Matrices In the following, we let (V, F) and (W, F) be vector spaces over the same field and we let ^ : y -^ W denote a linear mapping. We let {v^ . . . , v"} be a basis for V and
we set v^ = ^ v ^ . . . , v" = siv^. Then it is an easy matter to show that if v is any vector in V and if ( a i , . . . , a^) are the coordinates of v with respect to {v^ . . . , v"}, then ^ v = aiv^ + • • • + a„v". Indeed, we have siv = si(aiv^ + • • • + a^v") = ai^v^ 4- • • • + a„64v" = aiv^ + • • • + a„v". Next, we let {v^ . . . , v'^} be any set of vectors in W. Then it can be shown that there exists a unique Hnear transformation si from V into W such that ^v^ = v^ . . . , ^^v'^ = v". To show this, we first observe that for each v E V we have unique scalars ai,.. .,an such that V = aiv^ + ••• + a„v". Now define a mapping M : V ^
W as
1,.. .,n. We first must show that d- is Hnear and, then, that Clearly, ^ ( v ^ = v^ / aiv^ + • • • + a„v" and w = /3iv^ + • • • + jS^v'^, we have ^ is unique. Given v ^ ( v + w) = 6 4 [ ( a i + ) 8 i y + --- + (a^ + i 8 > " ] = (aj+/3i)v^ + •••+ (a„ + iS^)v^ On the other hand, d(v) = aiv^ + • • • + a^v'^, d(w) = /3iv^ + • • • + j8„v". Thus, ^(v) + ^(w) = (aiyi + • • • + a^v^) + (jSiv^ + • • • + ^^v^) = (^i + f3i)v^ + • • • + (a„ + i8„)v" = 62i(v + w). In a similar manner, it is easily established that ad(v) = d(av) for all Q: G F and v G V. Therefore, ^ is linear. Finally, to show that ^ is unique, suppose there exists a linear transformation ^ : V ^ W such that ^v^ = v^ / = 1 , . . . , /2. It follows that (d - ^)v^ = 0, / = 1 , . . . , n, and, therefore, that These results show that a linear transformation is completely determined by knowing how it transforms the basis vectors in its domain, and that this linear transformation is uniquely determined in this way. These results enable us to represent linear transformations defined on finite-dimensional spaces in an unambiguous way by means of matrices. We will use this fact in the following. Let (V, F) and {W, F) denote n-dimensional and m-dimensional vector spaces, respectively and let {v^ . . . , v^} and {w\ . . . , w^} be bases for V and W, respectively. Let V ^ W be a linear transformation and let v' = siv\ / = 1, n. Since {w^ w^} is a basis for W, there are unique scalars {aij}, i = I,.. .,m, j 1 , . . . , n, such that siv
= V = aiiw
+ a2iw + • • • + a^iw^
siv^ = Sp' = a\2W^ 4- a22>v^ + • • • + a^2>^^
(2.25)
^v^ = y" = aiyiW^ + a2nW^ + • • • + amn"^^' Next, let V G V. Then v has the unique representation v = aiv^ + 0:2v^ H h a^v" with respect to the basis {v^ . . . , v"}. In view of the result given at the beginning of this subsection, we now have ^ v = aiv^ + • • • + anv"".
(2.26)
Since ^^v G W, d-v has a unique representation with respect to the basis {w^ . . . , w^}, say, ^ v = jiw^ + 72W^ + • • • + 7mw'^.
(2.27)
105 CHAPTER 2 :
Response of Linear Systems
106 Linear Systems
Combining (2.25) and (2.26), we have ^^, - rv (n ^A^ -u
4. ^ ^^^^\
+ ^2(^121^^ + • • • + Clm2^^) + aniainW^
+ ••• +
amn^"^)'
Rearranging this expression, we have ^ v = (cin^i + ^i2<^2 + • • • + ai^a„)w^ + (^21 a; 1 + a22<^2 + • * * + a2nOin)'^^ + (amiai
+ am20^2 + • • • +
amnO^n)'^'^'
In view of the uniqueness of the representation in (2.27), we have 71 = a\\a\ + ^120:2 + • • • + ainOLn 7 2 == Cl2\0i-\ + <322«2 + • • • + a2^Q^Az
(2.28)
Tm = ^ m l « l + Clm20^2 + * * • + ^m^Q^m
where ( a i , . . . , anf and ( 7 1 , . . . , 7m)^ are coordinate representations of v E V and siv ^ W with respect to the bases {v^ . . . , v"} of V and {w^,..., w^} of W, respectively. This set of equations enables us to represent the linear transformation si from the linear space V into the linear space W by the unique scalars {atj}, i = I,.. .,m, j = ! , . . . , / ! . For convenience we let [atj]
an
ai2
...
CI21
^22
• ••
ain ^2n
^m\
^m2
'• •
^mn
(2.29)
We see that once the bases {v^ . . . , v^}, {w^ . . . , w^} are fixed, we can represent the linear transformation d- by the array of scalars in (2.29) that are uniquely determined by (2.25). Note that thejth column of A is the coordinate representation of the vector Av^ G W with respect to the basis [w^,..., w^}. In view of the results given at the beginning of this subsection, the converse to the preceding statement also holds. Specifically, with the bases for V and W still fixed, the array given in (2.29) is uniquely associated with the linear transformation ^ o f Vinto W. The above discussion gives rise to the following important definition. DEFINITION 2.3. The array given in (2.29) is called the matrix A of the linear transformation si from a linear space V into a linear space W (over F) with respect to the basis {v^ ..., v"} of V and the basis {w\ ..., w^} of W. • If in Definition 2.3, V = W, and if for both V and W the same basis {v^ . . . , v'^} is used, then we simply speak of the matrix A of the linear transformation si with respect to the basis {v^ . . . , v"}. In (2.29) the scalars (an, ai2,..., atn) form the /th row of A and the scalars (aij, a2j,..., amj)^ form thejth column of A. The scalar atj refers to that element of
matrix A that can be found in the ith row and jth column of A. The array in (2.29) is said to be an m X n matrix. If m = n, we speak of a square matrix. Consistent with the above, an ^ X 1 matrix is called a column vector, column matrix, or n-vector, and a 1 X n matrix is called a row vector. Finally, if A === [a/y] and B = [btj] are two mX n matrices, then A = B, i.e., A and B are equal if and only if aij = btj for all / = 1 , . . . , m, and for all j = 1 , . . . , n. Furthermore, we call A^ = [ciijV = l^ji] the transpose of A. The preceding discussion shows in particular that if ^ is a linear transformation of an n-dimensional vector space V into an m-dimensional vector space W, (2.30)
W == SiVy
if 7 = ( y i , . . . , ymV denotes the coordinate representation of w with respect to the basis {w^,..., w^}, if a = ( ^ i , . . . , a„)^ denotes the coordinate representation of v with respect to the basis {v\ . . . , v"}, and if A denotes the matrix of ^ with respect to the bases {v^ . . . , v"}, {w^ . . . , w"^}, then 7 = Aa,
(2.31)
or equivalently. = 1,... , m,
(2.32)
which are alternative ways to write (2.28). Some important remarks 1. Throughout this section we use, in the interests of clarity of presentation, lowercase Greek letters to denote the coordinate representations of vectors [see (2.31)]. In the interests of simplicity, however, we will use common (Latin) lowercase letters to denote vectors throughout the remainder of this book, whether they are coordinate representations of vectors or underlying objects (elements ofV). 2. We note that if, in particular, V = R^, then v E.V and its coordinate representation 7] with respect to the natural basis {^^ . . . , ^^} of V will have the same form.
*F. Some Properties of Matrices The rank of a matrix We first consider the characterization of the rank of a hnear transformation in terms of its matrix representation. To this end, let ^ be a linear transformation from a vector space V into a vector space W. It is easily shown that d- has rank r if and only if it is possible to choose a basis {v^ . . . , v"} for V and a basis {w^ . . . , w^} for W such that the matrix A of ^ with respect to these bases is of the form
107 CHAPTER 2 :
Response of Linear Systems
108 Linear Systems
rioo ••
A =
010
••
000 • •• 000 • ••
000 000
•• ••
100 • •• 000 • ••
° 0
000
••
000 • ••
0J
oT 0
> m = dim W.
Y n = dim V
More directly, if A is the matrix representation of ^ E L(V, W) with respect to some arbitrary bases {v^ . . . , v"} and {w^,..., w'"}, then (i) the rank of ^ is the number of vectors in the largest possible linearly independent set of columns of A; and (ii) the rank of ^ is the number of vectors in the smallest possible set of columns of A that has the property that all columns not in it can be expressed as linear combinations of the columns in it. The above result enables us now to make the following definition: the rank of an mX n matrix A is the largest number of linearly independent columns of A. General properties of matrices Since matrices are representations of linear transformations on finite-dimensional vector spaces (in the sense defined in Subsection E of this section), it is reasonable to suspect that matrices inherit the properties of the transformations they represent. In the following, we address this issue. Let V and W be ^-dimensional and m-dimensional vector spaces over F, respectively, and let ^ and ^ be linear transformations of V into W. Let A = [atj] and B = [bij] be the matrix representation of si and 2^, respectively, with respect to the bases {v^ . . . , v"} in V and {w^ . . . , w^} in W. Using (2.2) and Definition 2.3, the reader can readily verify that the matrix of the linear transformation ^ + 2S (with respect to the above bases), is given by A + B = [atj] + [bij] = [atj + btj] = [dj] = C.
(2.33)
Also, using (2.3) and Definition 2.3, the reader can easily show that the matrix of a A, denoted by D = a A, is given as a A -= a[aij] = [aatj] = [dfj] = D.
(2.34)
From (2.33) we note that for two matrices A and B to be added, they must have the same number of rows and columns. When this is the case, we say that A and B are comparable matrices. Clearly, if A is an m X n matrix, then so is aA. Next, let U be an r-dimensional vector space, let si E L(K W), and let 2^ G L{W, U). Let A be the matrix of d^ with respect to the basis {v^ . . . , v"} in V and with respect to the basis {w^,..., w^} in W. Let SS be the matrix of B with respect to the basis {w^ . . . , w^} in W and with respect to the basis {u^,.. .,u^}m U. The product mapping SS^ as defined by (2.6) is a linear transformation of V into U. By applying definitions, it is readily verified that the matrix of ^^ with respect to the
bases {v^ .. .,v^}ofVand
109
{u^,.. .,u^}of Uis given by
where
C = [Cij] = BA,
(2.35)
Cij = ^
(2.36)
bikakj
k=i
for / = 1 , . . . , r, and j = I,.. .,n. Clearly, two matrices A and B can be multiplied to form the product BA if and only if the number of columns of B is equal to the number of rows of A. When this is true, we say that the matrices B and A are conformal matrices. As mentioned earlier, the properties of general transformations established in Subsection D hold of course in the case of their matrix representations as well. We summarize some of these in the following: 1. Let A and BhemX
n matrices, and let C be an n X r matrix; then (A + B)C = AC + BC.
2. Let A be an m X n matrix, and let B and ChtnXr
(2.37) matrices; then
A{B + C) = AB + AC.
(2.38)
3. Let A be an m X n matrix, let 5 be an ^ X r matrix, and let C be an r X 5" matrix; then A{BC) = (AB)C
(2.39)
4. Let a, (3 E F , and let A be an m X n matrix; then (a + /3)A = aA + j8A.
(2.40)
5. Let a G F , and let A and BhQ mX n matrices; then a(A + B) = aA + aB.
(2.41)
6. Let a, (3 E: F, let A be an m X n matrix, and let 5 be an n X r matrix; then (aAXfiB)
= (aPXAB).
(2.42)
7. Let A and Bbe mX n matrices; then A + B = B + A. 8. Let A, B, and ChtmXn
(2.43)
matrices; then {A + B) + C = A + {B-^C).
(2.44)
Next, let 0 G L{V, W) be the zero transformation defined by (2.4). Then for any bases {v^ . . . , v"} and {w^ . . . , w^} for V and W, respectively, the zero transformation is represented by the m X n matrix
Too
•••
0
O = 00
•••
0
00
(2.45)
0
called the null matrix. Further, let 3 E: L(V, V) be the identity transformation defined by (2.12) and let {v\ . . . , v'^} be an arbitrary basis for V. Then the matrix
CHAPTER 2 :
Response of Linear Systems
110 Linear Systems
representation of the linear transformation 3" from V into V with respect to the basis {v\ • • •, v«} is given by ' 1 0 0 ••• 0 (2.46) / = 0 1 0 ••• 0 000
•••
1
called the n X n identity matrix. For any mX n matrix A we have that A^O
= 0 + A = A,
(2.47)
and for any nX n matrix B we have BI = IB = B, (2.48) where / is the n X n identity matrix. If A = [uij] is a matrix representation of a linear transformation ^4, then it is easily verified that the matrix -A == ( - l ) A = [-a/y] is the corresponding matrix representation of the linear transformation -s^. In this case it follows immediately that A -\- (-A) = O, where O denotes the null matrix. By convention we write that
A + (-A) =
A-A.
Next, let A and Bhe nX n matrices. Then we have in general that AB 7^ BA,
(2.49)
as was the case in (2.11). Further, let ^ G L(V, V), and assume that si is nonsingular with inverse ^ ~ ^ so that MM~^ = si~^si = ^. Now if A is the n X n matrix of si with respect to the basis {v^ . . . , v"} in V, then there is an n X ^ matrix B of M~^ with respect to the basis {v\ . . . , v'^} in y such that BA = AB = I.
(2.50)
We call B the inverse of A and we denote it by A~ ^ Under the present circumstances we say that A~^ exists, or A has an inverse, or A is invertible, or A is nonsingular. If A"^ does not exist, we say that A is singular From corresponding properties given in Subsection D for arbitrary linear transformations, several properties of matrices are evident. In particular, for an n X n matrix, the following are equivalent: (i) rank A — n\ (ii) Aa = 0 implies a = 0; (iii) for every yo ^ F^, there is a unique ao E F^ such that yo = Aao; (iv) the columns of A are linearly independent; and (v) A~^ exists. In Subsection E it was shown that we can represent n linear equations by the matrix equation (2.32) y = Aa.
(2.51)
Now assume that A is nonsingular. If we premultiply both sides of this equation by A~^ we obtain a = A-^y,
(2.52)
the solution to Eq. (2.51). Thus, knowledge of the inverse of A enables us to solve the system of linear equations (2.51). Concerning inverses of matrices, the following facts are easily verified: (i) an nXn nonsingular matrix has one and only one inverse; (ii) if A and B are nonsingular
nX n matrices, then (AB)~^ = B~^A~^', and (iii) if A and B diVQnX n matrices and if AB is nonsingular, then so are A and B. Next, we consider the principal properties of the transpose of matrices, which follow readily from definitions: (i) for any matrix A, (A^)^ = A; (ii) if A and B are conformal matrices, then (ABY = B^A^; (iii) if A is a nonsingular matrix, then (A^)"^ = (A~^)^', (iv) if A is an n X n matrix, then A^ is nonsingular if and only if A is nonsingular; (v) if A and B are comparable matrices, then (A + B)^ = A^ + B^; and (vi) if a G F and A is a matrix, then (aA)^ = aA^. Next, we let A he an n X n matrix, and we let m be a positive integer. As in (2.16), we define the nX n matrix A^ by A^ = /{'A
4,
(2.53)
"Y" m times
and if A ^ exists, then as in (2.17) we define the nX n matrix A '^ as A-^
l\m = (A-'y
A
•• -A"
(2.54)
V m times
As in the case of Eqs. (2.18) to (2.20), the usual laws of exponents follow from the above definitions. Specifically, if A is an n X n matrix and if r and s are positive integers, then A'--A' (Ay
= A'^' = A'^' = A' • A' = A " = A'' = (A'Y,
(2.55) (2.56)
A' • A-' =
(2.57)
and if A"' exists, then A'-'.
Consistent with this notation, we have A' = A lO
and
^
(2.58) (2.59)
We are now once more in a position to consider functions of linear transformations, where in the present case the linear transformations are represented by matrices. For example, if /(A) is the polynomial in A given in (2.23), and if A is any nXn matrix, then by /(A) we mean /(A) = a o / + aiA + • • • + a^A^.
(2.60)
Finally, we noted earlier that, in general, linear transformations (and in particular, matrices) do not commute [see (2.11) and (2.49)]. However, in the case of square matrices, the following facts are easily verified: let A, B, C denote nX n matrices; let O denote the nX n null matrix; and let / denote the n X n identity matrix. Then (i) O commutes with any A; (ii) A^ commutes with A^, where/? and q are positive integers; (iii) a I commutes with any A, where a E: F', and (iv) if A commutes with B and if A commutes with C, then A commutes with aB -\- pC, where a, (3 E F . *G. Determinants of Matrices Definition of determinant In this section we recall and summarize some of the important properties of determinants of a matrix. To this end we let N = {1,2,.. .,n} and recall that a
111 CHAPTER 2 :
Response of Linear Systems
112 Linear Systems
permutation on A/^ is a one-to-one mapping ofN onto itself. For example, if O" denotes a permutation on N, then we can represent it symbolically as 1 2--where ji G A/^ for / = 1, (7 more compactly as
n
(2.61)
, n, and jr ^ jk whenever r ^ k. Henceforth, we represent (y =
JlJ2---jn-
Clearly, there are nl possible permutations on N. We let P{N) denote the set of all permutations on N, and we distinguish between odd and even permutations. Specifically, if there is an even number of pairs (/, k) such that / > k but / precedes k in O", then we say that a is even. Otherwise, a is said to be odd. Finally, we define the function sgn from P{N) into F by
+1,
sgn (cj)
if (7 is even if a is odd
forallcJGP(A^). As a specific example, foxN= { 1 , 2 , 3 } , there are six different permutations, even and odd, on N given in the following table:
G
Uuh)
123 132 213 231 312 321
(1,2) (1,3) (2,1) (2,3) (3,1) (3,2)
Uuh)
U2J3)
G is odd or even
(1,3) (1,2) (2,3) (2,1) (3,2) (3,1)
(2,3) (3,2) (1,3) (3,1) (1,2) (2,1)
Even Odd Odd Even Even Odd
Now let A denote the nxn
sgn(G)
-1
matrix given by 'an ail
an ail
a\n ain
\_afii
a^ii
ann.
We form the product of n elements from A by taking one and only one element from each row and one and only one element from each column. We represent this product as ^Ijl '^2J2
anj^^
where {jiji • • • jn) ^ ^ ( ^ ) - H is possible to find n\ such products, one for each a G P{N). We are now in a position to define the determinant of A, denoted by det{A), by the sum det{A) =
^ sgn{G) • aij^ • ay^ oePiN)
^njn^
(2.62)
where a = jvin-
We frequently denote the determinant of A by
113 CHAPTER 2 :
Cl\n Clin
det{A) =
= \A[
(2.63)
Properties of determinants We now enumerate some of the common properties of determinants. The proofs of these follow mostly from definitions. Let A and BhonXn matrices. Then (i) det (A^) = det (A); (ii) if all elements of a column (or row) of A are zero, then det (A) = 0; (iii) if the matrix B is the matrix obtained by multiplying every element in a column (or row) of A by a constant a, while all other columns of B are the same as those of A, then det(B) = a det{A)\ (iv) if JB is the same as A, except that two columns (or rows) are interchanged, then det(B) == -J^f (A); (v) if two columns (or rows) of A are identical, then J^^( A) = 0; and (vi) if the columns (or rows) of A are linearly dependent, then det (A) = 0. We now introduce some additional concepts for determinants. To this end, let A = [aij] be an /I X n matrix. If the /th row andjth column of A are deleted, the remaining (n - 1) rows and (n - 1) columns can be used to form another matrix Mij whose determinant is det (Mij). We call det {Mij) the minor of aij. If the diagonal elementsofM/y are diagonal elements of A, i.e., if / = j , then we speak of aprmc/pa/ minor of A. The cofactor of aij is defined as ( - ly^^det (Mij). As a specific example, if A is a 3 X 3 matrix, then det (A) =
an
ai2
au
^21
^22
Cl23
^31
^32
<^33
the minor of element ^23 is det(M23) = an
ai2
«31
CI31
and the cofactor of ^23 is CTh
a\\ (-1) <23i
a\2 (332
Next, for an arbitrary nXn matrix A, let c/y denote the cofactor of aij, i, j = I,.. .,n.lt can be shown from definitions that the determinant of A is equal to the sum of the products of the elements of any column (or row) of A, each by its cofactor. Specifically, det (A) = ^
^ijCij,
j = \,..
.,n,
(2.64)
i=\
or
det (A) = ^^aijCij,
i = 1,..
For example, if A is a 2 X 2 matrix, we have det (A)
an
a\2
<^2i
^22
-
<2ll^22 "
^12^2b
(2.65)
Response of Linear Systems
114
and if A is a 3 X 3 matrix, we have
Linear Systems det{A)
kll = \a2i
1^31
^12 ^22
^32
^13 <^23
^33!
= an
Cl22
<^23
^32
(233
-
^12
(221
^23
Cl2\
'^22
^31
<^33
+ a\2> (331
^32
= ancn + ^21^21 -^ cisic^i. We now consider a few additional useful properties of determinants. In particular from basic definitions it can be shown that if the ith row of an n X ^ matrix A consists of elements of the form an + a'-p ai2 + a[2,.. .,ain + a\^, i.e., if
A =
an
(212
^1«
<221
<322
^2n
(an + a'.^)
((2,-2 + ^;2)
(atn + (2;^)
<^/22
an
a\2 ^2n
then
det(A) =
<321
<322
<3/l
(2/2
<2«1
^n2
an
a\2
Cl2\
C122
a\^
^a
(^n\
Cln2
(^2n
Next, if A and B axe n X n matrices, and if B is obtained from A by adding a constant a times any column (or row) to any other column (or row) of A, then it is easily shown that (i^r (A) = det(B). Further, if ctj denotes the cofactor of atj, i, j = 1 , . . . , n, for annX n matrix A, then it is easily shown that ^atjCik
=- 0
foxJT^k
(2.66)
= 0
for / 7^ k.
(2.67)
i=\
and
^^CLijCkj 7=1
We can combine (2.64) with (2.66) and (2.65) with (2.67) to obtain the relations n
^atjCij^
= det(A)Sjk,
j , k = \,...,n,
= det(A)81^,
i, k = \,..
(2.68)
/= i
and
^atjCkj
.,n,
(2.69)
7=1
respectively, where 8mn denotes the Kronecker delta (i.e., 8mn = 1 when m = n and ^mn = 0 Otherwise). An extremely useful result concerning determinants (which can be proved by using the definition of determinant and some of the properties enumerated above) states that the determinant of the product of two matrices is equal to the product of the determinants of the matrices. Thus, if A and B are nX n matrices, then det(AB) =
det(A)det(B).
(2.70)
It is easily verified that for the n X n identity matrix / and for the n X n zero matrix O, we have det (I) = 1 and det (O) = 0. Finally, we can readily show that annX n matrix A is nonsingular if and only if det (A) ¥=0. We conclude this discussion by considering a means of determining the inverse, A~^ of a nonsingular nX n matrix. To this end, let c/y be the cofactor of atj, i, j = 1 , . . . , n, for the matrix A, and let C be the matrix formed by the cofactors of A, i.e., C = [cij]. The matrix C^ is called the (classical) adjoint of A, and is denoted by adj (A). It is easily verified that A • [adj (A)] = [adj (A)] • A - [det (A)] • /,
(2.71)
from which it follows that 1 adj (A). det (A)
(2.72)
As a specific case, consider 0 1 1
A =
1 2 -1
1 2 0
Then det (A) - - 1 , adj (A) =
and
A"
2 2 3
-1 0' -1 1 1 -1
2 1 0" 2 1 -1 3 -1 1
H. Solving Linear Algebraic Equations Now consider the linear system of equations given by Aa = y,
(2.73)
where A E R^^^ and y E. R^ are given and a G /?" is to be determined. By using the results of the preceding subsections, especially Subsection 2.2D, the following important results can readily be established. 1. For a given y, a solution a of (2.73) exists (not necessarily unique) if and only if 7 G 2^(A), or equivalently, if and only if p{[A,y^) = p{Ar
(in A)
2. A solution a of (2.73) exists for any y if and only if p(A) = m.
(2.75)
If (2.75) is satisfied, a solution of (2.73) can be found by using the relation a = A'^(AA^y^y.
(2.76)
115 CHAPTER 2 :
Response of Linear Systems
116 Linear Systems
When in (2.73), p(A) = m = n, then A G j^nxn ^^^ jg nonsingular and the unique solution of (2.76) is given by a = A~^y,
(2.77)
3. Every solution a of (2.73) can be expressed as a sum a = ap-\- ah,
(2.78)
where ap is a specific solution of (2.73) and ah satisfies Aah = 0. This result allows us to span the space of all solutions of (2.73). Note that there are dim>f(A) = n-
p{A)
(2.79)
linearly independent solutions of the system of equations AjS = 0. EXAMPLE 2.7. Consider 0
0
Aa = 0 0 .0
0
0" 1 0.
(2.80)
r-
It is easily verified that {(0, 1, 0)^} is a basis for S/l(A). Since a solution of (2.80) exists if and only if y G ^(A), y must be of the form y = (0, k,0), k G R. Note that 0
0
0
p(A) = I = p([A, y]) = rank 0 0
1
k
0
0
0
0
0
as expected. To determine all solutions of (2.80), we need to determine an a^ and an ah [see (2.78)]. In particular, ap = (0,0, Z:)^ will do. To determine an, we consider AjS = 0. There are dim >r(A) = 2 linearly independent solutions of A/3 = 0. In particular, {(1, 0, 0)^, (0, 1, 0)^} is a basis for M(A). Therefore, any solution of (2.80) can be expressed as ap -\- ah
0' 0 .k.
T
01
+ 0 1 \ci .0 oj l_<^2.
where ci, C2 are appropriately chosen real numbers. We conclude by noting that an extensive discussion of determining solutions of the linear system of equations (2.73) is provided in the Appendix.
I. Equivalence and Similarity From our previous discussion it is clear that a linear transformation ^4 of a finitedimensional vector space V into a finite-dimensional vector space W can be represented by means of different matrices, depending on the particular choice of bases in V and W. The choice of bases may in different cases result in matrices that are easy or hard to utilize. Many of the resulting "standard" forms of matrices, called canonical forms, arise because of practical considerations. Such canonical forms often exhibit inherent characteristics of the underlying transformation si. Throughout the present subsection, V and W are finite-dimensional vector spaces over the same field F, dim V = n, and dim W = m.
•
Change of bases: Vector case Our first aim will be to consider the change of bases in the coordinate representation of vectors. Let {v^ . . . , v""} be a basis for V and let {v^ . . . , v'^} be a set of vectors in V given by
=
(2.81)
i = \,..
^Pn
where pij G F for all i, j = 1 , . . . , n. It is easily verified that the set {v^ . . . , v"} forms a basis for V if and only if the n X n matrix P = [pij] is nonsingular. We call P the matrix of the basis {v\ . . . , v"} with respect to the basis {v^ . . . , v"}. Note that the /th column of P is the coordinate representation of v' with respect to the basis Continuing the above discussion, let {v^ . . . , v"} and {v\ . . . , v'^} be two bases for y, and let P be the matrix of the basis {v^ . . . , v"} with respect to the basis {v^ . . . , v"}. Then it is easily shown that P~^ is the matrix of the basis {v^ . . . , v'^} with respect to the basis {v^ . . . , v"}. Next, let the sets of vectors {v^ . . . , v"}, {v^ . . . , v^}, and {v^ . . . , v^} be bases for V. If P is the matrix of the basis {v^ . . . , v"} with respect to the basis {v^ . . . , v"} and if Q is the matrix of the basis {v V • •, v"} with respect to the basis {v^ . . . , v'^}, then it is easily verified that PQ is the matrix of the basis {v^ . . . , v"} with respect to the basis {v^ . . . , v^}. Continuing further, let {v^ . . . , v"} and {v^ . . . , v"} be two bases for V and let P be the matrix of the basis {v^ . . . , v"} with respect to the basis {v^ . . . , v'^}. Let a ^ V and let a^ = ( a i , . . . , a„) denote the coordinate representation of a with respect to the basis {v^ . . . , v"} (i.e., a = XJ^= i cav^)- Let d^ = (di,..., dn) denote the coordinate representation of a with respect to the basis {v^ . . . , v"}. Then it is readily verified that Pd = a. EXAMPLE 2.8. Let V = R^ md F = R, and let a = (1, 2, 3)^ ( R^ be given. Let {v\ v^, v-^} = {e^, e^, e^} denote the natural basis for R^, i.e., e^ = (1,0, Of, ^2 ^ (0, 1, 0)^, e^ = (0,0,1)^. Clearly, the coordinate representation a of a with respect to the natural basis is (1, 2, 3)^. Now let {v\ v^, v^} be another basis for R^, given by v^ = (1, 0, 1)^, v2 = (0, 1, 0)^, = (0, 1, l)'^. From the relation (1,0,1)^ = v^ =
P\\V^ + PllV^ + P3lV^
rii
ro"
roi
0
1
= Pii 0 + P21 1 + P31 0
0
we conclude that pn = 1, p2i = 0, and psi = 1. Similarly, from (0, 1, 0)^ = v'^ = pnV^ + P22V^ + P32V^
\l~
roi
roi
0
1
P\2 0 + Pll 1 + P32 0
0
117 CHAPTER 2:
Response of Linear Systems
118 Linear Systems
we conclude that pu = 0, P22 = 1, and P32 = 0. Finally, from the relation rii roi roi (0,1,1)^ Pl3 0 + P23 1 + P33 0 0
0
1
we obtain that pu = 0, P23 = 1, and P33 = 1. The matrix P = [ptj] of the basis {v\ v^, v^} with respect to the basis {v^ v^, v^} is therefore determined to be ri 0 01 p = \o 1 1 , [1 0 ij and the coordinate representation of a with respect to the basis {v^ v^, v^} is given by d = P~^a, or ri 0 01 - 1 r 11 0 1 1 r .1 0 1, L:3j r 1 0 01 rr Ml 1 1 - 1 2 = 0 .-1 0 oj L3. _2_ Change of bases: Matrix case Having addressed the relationship between the coordinate representations of a given vector with respect to different bases, we next consider the relationship between the matrix representations of a given linear transformation relative to different bases. To this end let ^ G L(V, W) and let {v^ . . . , v^} and {w^ . . . , w"^} be bases for V and W, respectively. Let A be the matrix of si with respect to the bases {v^ . . . , v"} and {w^,..., w"^}. Let {v^ . . . , v^} be another basis for V, and let the matrix of {v^ . . . , v'^} with respect to {v^ . . . , v'^} be P. Let {w^,..., w^] be another basis for W, and let Q be the matrix of {w^ . . . , w^} with respect to {vP^ . . . , w^}. Let A be the matrix of ^ with respect to the bases {v^ . . . , v'^} and {w^ . . . , w^}. Then it is readily verified that A = QAR
(2.82)
This result is depicted schematically in Fig. 2.1.
V-^ W V = PV
A —*
Pt {v\ • • •. v"} V
0) = Av \ Q
A cb = Qo)
FIGURE 2.1 Schematic diagram of the equivalence of two matrices
Equivalence of matrices The preceding discussion motivates the following definition.
DEFINITION 2.4. Anm X n matrix A is said to be equivalent to an m X ^ matrix A if there exists an m X m nonsingular matrix Q and an /i x ^ nonsingular matrix P such that (2.82) is true. If A is equivalent to A, we write A ~ A. • Thus, an m X ^ matrix A is equivalent to an m X n matrix A if and only if A and A can be interpreted as both being matrices of the same linear transformation ^4 of a linear space V into a linear space W, but v^ith respect to possibly different choices of bases. It is clear that a matrix A is always equivalent to itself (i.e., A ~ A). Also, if a matrix A is equivalent to a matrix 5, then clearly B is equivalent to A (i.e., if A ~ 5, then 5 — A). Furthermore, if A is equivalent to B and B is equivalent to a matrix C, then it is evident that A is equivalent to C (i.e., if A ~ 5 and B — C, then A ~ C). This shows that ~ is an equivalence relation. The reader can easily verify that every m X n matrix is equivalent to a matrix of the form
"100 •• 010 ••
•• ••
0 " 0 } r = rank A 0 0
000 000
•• ••
100 000
•• ••
000
••
000
••
0-
From this it follows that two mX n matrices A and B are equivalent if and only if they have the same rank, and furthermore, that A and A^ have the same rank. The definition of rank of a matrix that we used in Section 2.2 is sometimes called the column rank of a matrix. Sometimes, an analogous definition for row rank of a matrix is also used. The result given in the above paragraph shows that row rank of a matrix is equal to its column rank. Similarity of matrices Next, let V = W, let 5i e L{V, V), let {v^,..., v"} be a basis for V, and let A be the matrix of d. with respect to {v^ . . . , v"}. Let { v ' , . . . , v"} be another basis for V whose matrix with respect to {v^ . . . , v"} is P. Let A be the matrix of ai with respect to {v^ . . . , v"}. Then it follows immediately from (2.82) that A = P'^AP.
(2.83)
The meaning of this result is depicted schematically in Fig. 2.2. This discussion motivates the following definition.
V {v\...,v"}
{v\...,v"}
t P
IP-'
{v , . . . , v"}
A -^
{v ,..., v"}
FIGURE 2.2 Schematic diagram of the similarity of two matrices
119 CHAPTER 2 :
Response of Linear Systems
120 Linear Systems
DEFINITION 2.5. knnXn matrix A is said to be similar to em nXn matrix A if there exists SinnX n nonsingular matrix P such that A = P-^AP. If A is similar to A, we write A -- A. We call P a similarity transformation. • It is easily verified that if A is similar to A [i.e., (2.83) is true], then A is similar to A, i.e., A = PAp-\ (2.84) In view of this, there is no ambiguity in saying "two matrices are similar," and we could just as well have used (2.84) [in place of (2.83)] to define similarity of matrices. To sum up, if two matrices A and A represent the same linear transformation d^ G L{V, y), possibly with respect to two different bases for V, then A and A are similar matrices. Since the similarity of two matrices is a special case of the equivalence of matrices, it follows that (i) a matrix A is similar to A; (ii) if A is similar to a matrix B, then B is similar to A; and (iii) if A is similar to B and B is similar to a matrix C, then A is similar to C. Therefore, the similarity relation of matrices is an equivalence relation. Now let A be an n X fz matrix that is similar to a matrix B. Then it is easily shown that A^ is similar to B^, where k denotes a positive integer, i.e., B^ = P~^A^P. This can be extended further by letting m
and by verifying that
ao + aiX + • • • + a^A"
(2.85)
i=0
f(p-'AP) = p-'f{A)P,
(2.86)
where a o , . . . , a ^ E F . This shows that if B is similar to A, then f{B) is similar to /(A), where in fact the same similarity transformation P is involved. Further, if A is similar to A and if /(A) is as given in (2.85), then /(A) = O if and only if /(A) == O. Next, let ^4 G L(K V) and let A be the matrix of ^4 with respect to a basis {v\ . . . , v"} in y. Let /(A) denote the polynomial given in (2.85) and let A be any matrix of si. Then it is readily verified that f{d) = 0 if and only if /(A) = O, We can use results such as the preceding to good advantage. For example, let A denote the diagonal matrix given by 'Ai 0 A =' 0 0 Then
\x\ 0
0 0 ••• A2 0 ••• 0 0
0 0
0 •• • A„-i 0 •• 0
0 0 0 An J
0 0 • A^ 0 •
0 0
0 0
0 0
0
0 AS
{Af = 0
0 • 0 •
Now let /(A) be given by (2.85). Then
[l 0
0 ••• . . . 1
•••
...
0 0
121 Ai 0
0
•
A2
•
0 0 |_0 0
••• •••
"AY* 0
0 A-
0 .0
0 0
1 0
0 1
Response of Linear Systems
+ •••
+ ai
f{A) = ao
CHAPTER 2 :
0 0
0
[0
0 - ••
A„-i
0
•
0
An
0
01 0
7(Ai) 0
0 /(Ai)
0 A"*
0 0
0 0
0 0
•
+ a„ 0
/(A„-i) 0
0 /(A„)
Next, let ^ E L{V, V), let A be the matrix of si with respect to a basis {v^ . . . , v"} in y, and let A be the matrix of si with respect to another basis {v^ . . . , v'^} in V. Then it is easily verified that det (A) = del (A). From this it follows that for any two similar matrices A and B, we have det (A) = det (B). In view of these results, there is no ambiguity in defining the determinant of a linear transformation 6^^ of a finite-dimensional vector space V into V as the determinant of any matrix A representing it, i.e., det(^) = det (A).
J. Eigenvalues and Eigenvectors Definitions Throughout this subsection, V denotes an n-dimensional vector space over a field F. Let ^ E L(V, V) and let us assume that there exist sets of vectors {v^ . . . , v"} and {v^ . . . , v"} that are bases for V such that =
= Aiv'
v2 =
=W
V
v« = siv" = A„v", where Xi ^ F,i = I,..., n. If this is the case, then the matrix Aof d. with respect to the given bases is [Ai
0
0
A„
A =
This motivates the following result that is easily verified: for M G LiV, V) and A G F, the set of all v G V such that siv = Av
(2.87)
122 Linear Systems
is a linear subspace of V. In fact, it is the null space of the linear transformation (^ ~~ ^^)^ where ^ is the identity element of L(V, V). Henceforth, we let JVTA = {v G y : ( ^ - Ai)v = 0}.
(2.88)
The above gives rise to several important concepts that we introduce in the following definition. DEFINITION 2.6. A scalar A such that Kx [given in (2.88)] contains more than just the zero vector is called an eigenvalue of ^4 (i.e., if there is a v T^ 0 such that siv = Av, then A is called an eigenvalue of si). When A is an eigenvalue of si, then each v T^ 0 in J^TA is called an eigenvector of ^ corresponding to the eigenvalue A. The dimension of the linear subspace JVA is called the (geometric) multiplicity of the eigenvalue A. If JVA is of dimension one, then A is called a simple eigenvalue. The set of all eigenvalues of d^ is called the spectrum of si. • Other names for eigenvalue that are in use include proper value, characteristic value, latent value, or secular value. Similarly, other names for eigenvectors are proper vector or characteristic vector. The space Mx is called the Ath proper subspace of V. For matrices, we give the following corresponding definition. DEFINITION 2.7. Let A be an « X ^ matrix whose elements belong to the field F. If there exists A E F and a nonzero vector a E F"^ such that Aa = Xa,
(2.89)
then A is called an eigenvalue of A and a is called an eigenvector of A corresponding to the eigenvalue A. • The connection between Definitions 2.6 and 2.7 is given in the following result that the reader can verify easily: let d^ E: L{V,V) and let A be the matrix of ^ with respect to the basis {v^ . . . , v"}. Then A is an eigenvalue of d^ if and only if A is an eigenvalue of A. Also, a G V is an eigenvector of si corresponding to A if and only if the coordinate representation of a with respect to the basis {v^ . . . , v"}, a, is an eigenvector of A corresponding to A. We note that if a (or a ) is an eigenvector of si (of A), then any nonzero multiple of a (of a ) is also an eigenvector of si (of A). In the case of matrices, in place of (2.89), one can also consider the relationship aA = Xa,
(2.90)
where a denotes a 1 X n row vector. In this context, a. in (2.89) and a in (2.90) are referred to as a right eigenvector and a left eigenvector, respectively. Unless explicitly stated, we will have in mind a right eigenvector when using the term eigenvector of a matrix. Now let si G L{V, V) and let A denote the matrix of si with respect to the basis {v^ . . . , v"} in y. Then it is easily shown that A E F is an eigenvalue of d^ (and hence, of A) if diViA only if det {si-X3>) = 0, orequivalently, if andonly if J ^ r ( A - A / ) = 0. Characteristic polynomial The above result enables us to determine the eigenvalues of d (or A) in a systematic manner. So let us examine the equation det{si-
X3) = 0
(2.91)
or equivalently, the equation
123
det{A- \I) = 0
(2.92)
in terms of the parameter A. We first rewrite (2.92) as 1(^11 - A)
det(A-\I)
=
(212
Clll
(Cl22 -
^nl
^n2
Clin A)
^2n
'"
(2.93)
(^nn ~ A)|
It is clear from (2.62) that the expansion of the determinant (2.93) yields a polynomial in A of degree n. For A to be an eigenvalue of si (or A) it must satisfy (2.91) [or (2.92)] and it must belong to F. Note that in general we have no assurance that the fzth-degree polynomial given by (2.92) has any roots in F. There is, however, a special class of fields for which this requirement is automatically satisfied: a field F is said to be algebraically closed if for every polynomial p(X) there is at least one A G F such that p{X) = 0.
(2.94)
Any A that satisfies (2.94) is said to be a root of the polynomial equation (2.94). In particular the field of complex numbers is algebraically closed, whereas the field of real numbers is not (e.g., consider the equation A^ + 1 = 0). There are other fields besides the field of complex numbers that are algebraically closed. However, since we will not require these, we will restrict ourselves to the field of complex numbers, C, whenever the algebraic closure property is required. When considering results that are valid for a vector space over an arbitrary field, we will, as before, make use of the symbol F, or make no reference to F at all. Summarizing the above discussion, we have the following result. Let ^ G L(V, V) and let A be the matrix of ^ with respect to the basis {v^ . . . , v"} in V. Then (i) det (^ - \3) = det (A - A/) is a polynomial of degree n in the parameter A, i.e., there exist scalars ao, a i , . . . , a„, depending on d^ (and therefore on A) such that det{d^ - X3) = det (A - A/) = ao + o^i^ + «2A^ + ••• + a^A'^
(2.95)
[note that ao = det (si) and an = (-l)'^]; (ii) the eigenvalues of ^ are precisely the roots of the equation det(M - A^) = det (A - A/) = ao + aiA + a2A^ + • • • + a^A'^ = 0;
(2.96)
and (iii) M^ has, at most, n distinct eigenvalues. We call (2.95) the characteristic polynomial of si (or of A) and call (2.96) the characteristic equation ofd (or of A). An important remark concerning notation The above definition of characteristic polynomial is the one usually used in texts on linear algebra and matrix theory (refer, e.g., to some of the books on this subject cited at the end of this chapter). An alternative to the above definition is given by the expression a(A) - det(X3 -A)
= det{XI - A).
One of the reasons for using this convention is that this polynomial arises in a natural manner when solving systems of ordinary differential equations by operator methods
CHAPTER 2 :
Response of Linear Systems
124 Linear Systems
[e.g., system (LH)]. Since the reader may have many occasions to consult texts on linear algebra and matrix theory, we will employ in this section the definition given in (2.95). Throughout the remainder of this book, however, we will follow the convention used in linear systems texts by utilizing the expression det (A/ - A) for the characteristic polynomial. We note that because of the relationship det (A - XI) = {-If
det {XI - AX
either definition can be used to develop the results considered herein. Note that det (XI - A) is a monic polynomial, i.e., its leading coefficient equals 1. From the fundamental properties of polynomials over the field of complex numbers, there now follows the next important result: if V is an n-dimensional vector space over C and if ^^i E LiV, V), then it is possible to write the characteristic polynomial of d^ in the form det (d - X3) = (Ai - A)^i(A2 - A)^^.. .(^^ _ x)'^p,
(2.97)
where Xu i = I,..., p, are the distinct roots of (2.96) (i.e.. A/ T^ XJ, if / T^ j). In (2.97), Mi is called the algebraic multiplicity of the root A/. The m/ are positive integers, and S f = i ^/ = ^• The reader should make note of the distinction between the concept of algebraic multiplicity of A^, given above, and the (geometric) multiplicity of an eigenvalue A/, given earlier. In general these need not be the same, as will be seen later. The Cayley-Hamilton Theorem and appUcations We now state and prove a result that is very important in linear systems theory. THEOREM 2.1. (CAYLEY-HAMILTON THEOREM) Every square matrix satisfies its characteristic equation. More specifically, if A is an n X n matrix and p(X) = det (A XI) is the characteristic polynomial of A, then p(A) = O. Proof. Let the characteristic polynomial for A be p{X) = ao + aiX-\ h a„A" and let B{X) = [bij(X)] be the classical adjoint of (A - XI) (refer to Subsection 2.2G). Since the bij(X) are cofactors of the matrix A - A/, they are polynomials in A of degree not more than n-l. Thus, bij(X) = Ptjo + ptjiX + ••• + Ptj^n-DX^'-K Letting Bj, = Wtjk] for A: = 0, 1,..., n - 1, we have B(X) = BQ + XBi + " + A^~^5„-i. By (2.71), we have (A~XI)B(X) = [det(A-XI)]LThus,(A-XI)[Bo + XBi-h"' + X''-^Bn-i] = (ao+Q:iA+ • • • + anX^)I. Expanding the left-hand side of this equation and equating like powers of A, we have -Bni = oLnL AB^i - Bn-2 = 0Ln-\I,..., AB\ - Bo = ail, ABQ = a^I. Premultiplying the above matrix equations by A", A"~\ ..., A, /, respectively, we have -A^Bn-i - Qf„A«,A«5„-i-A"-i5„-2 = a„-iA"-i, . . . , A 2 B I - A 5 O = a^AAB^ = aol. Adding these matrix equations, we obtain O = a^I + aiA + • • • + anA"^ = p(A), which was to be shown. • As an immediate consequence of the Cayley-Hamilton Theorem, we have the following results: let A be an n X n matrix with characteristic polynomial given by (2.96). Then (i) A^ - ( - l ) " + i [ « o / + ^lA + • • • + an-iA""-^]; and (ii) if/(A) is any polynomial in A, then there exist JSQ, jSi,..., /3„_i G F such that f(A)
= /3o/ + iSiA + • • • + /3n-iA'~\
(2.98)
Part (i) follows froir the Cayley-Hamilton Theorem and from the fact that an = (-1)". To prove part (ii), let /(A) be any polynomial in A and let p(X) denote the characteristic polynomial of A. From a result for polynomials (called the division
algorithm), we know that there exist two unique polynomials g{X) and r(A) such that
fil) = p{l)gil) + r{l),
(2.99)
where the degree of r(A)
1
t G {—a,a).
A^
(2.100)
k=0
In view of the Cayley-Hamilton Theorem, we can write /(A)
^At
(2.101) i=0
In the following, we present a method to determine the coefficients ai{t) in (2.101) [or/3/in (2.98)]. In accordance with (2.97), let p(A) = det{A - XI) = Uf=i{^i - ^T' be the characteristic polynomial of A. Also, let /(A) and g(A) be two analytic functions. Now if /«(A,0=g«(A,0, / = 0,...,m,--l,/=l,...,p, (2.102) where f^^\Xi) = ((i7/^^0(^)U=Aplf=i m = n, then /(A) = g{A). To see this, we note that condition (2.102) written as ( / - g)^^^ (A/) = 0 impUes that /(A) - g{X) has p(A) as a factor, i.e., /(A) — g(A) = w(A)p(A) for some analytic function w(A). From the Cayley-Hamilton Theorem we have that p{A) = O and therefore /(A) — g{A) = 0. -1 1 and let /(A) = e^\f{X) = e^\ and g{X) = aiA+ -1 1 OQ. The matrix A has an eigenvalue A = Ai = A2 = 0 with multiplicity m\ =2. Conditions (2.102) are given by /(Ai) = ^(Ai) = 1 and f^^\X\) = g^^\h) and imply that
E X A M P L E 2.9. Let A
125 CHAPTER 2 :
Response of Linear Systems
126 Linear Systems
ao = 1 andai = f. Therefore, ^A. ^ f^^^ ^ (^) ^ ^ ^ +
j ^ [-«! + «o
a^
] l l ~ t
t
In the final result of this section we let F = C, we let A be an n X n matrix of A G L(V; y), and we let det{A - A/) - det(A - \3) be given by (2.97). It is readily verified that (i) det(A) = [ l ^ . i ^J\ (ii) trace (A) = X^^iau = Z y = i m^Ay, (iii) if B is any matrix similar to A, then trace (B) = trace (A), and (iv) if /(A) denotes the polynomial /(A) = 7o + 7i A H \- 7mA^, then the roots of the characteristic polynomial of /(A) are / ( A i ) , . . . , /(A^), and J^^ [/(A) - A/] = [/(Ai) -
Ar---[/(A;,)-An. K. Direct Sums of Linear Subspaces One of the important topics in matrix theory is the development of canonical forms, including the lower (upper) triangular form, the block diagonal form, and the Jordan canonical form of matrices. Before presenting these, we need to address additional topics which are of interest and importance to us in their own right. One of these concerns direct sums of linear subspaces and linear transformations defined on such sums. Let y be a linear space and let W and U be arbitrary subsets of V. The sum of sets W and U, denoted by W -\- U, is the set of all vectors in V that are of the form w + w, where w E. W and u E U. If in particular, W and U are linear subspaces of y, then it is easily shown that W -\- U is also a linear subspace. If W and U are linear subspaces of V, and ifWnU = {0} (the singleton set consisting of the null vector of V), then we say that W and U are disjoint. Note that this terminology is not consistent with that used in connection with sets. If W and U are linear subspaces of y, then it is easily verified that for every v E U + W, there exist unique elements w ELW and u E U such that v = u -\- wif and only ifUHW = {0}. The preceding discussion is readily extended to any number of linear subspaces of y and gives rise to the following concept. Let y i , . . . , y^ be linear subspaces of a vector space V. The sum yi + • • • + y^ is said to be a direct sum if for each v G Vi -\ + Vr there is a unique set v^ G Vi, i = 1 , . . . , r, such that v = v^ + • • • + v'*. Henceforth, we will denote the direct sum of y i , . . . , y^ by yi © • • • © y^. Now let y be the direct sum of linear spaces Vi and V2, i.e., y =^ yi © V2, and let V = v^ -h v^ be the unique representation of v G y, where v^ G yi and v^ G y2. We say that the projection on Vi along V2 is the transformation defined by 2^(v) = v^
(2.103)
We can easily verify that (i) ^ G L(y y), (ii) 31(2^) - Vu and (ii) J{{^) = V2. More generally, we can show that if 2P G L(V, V), then 2^ is a projection on 9l(SP) along X(&) if and only if 2^2P = ?P^ = ?P. This gives rise to the following concept: ^ G L(V; V) is said to be idempotent if 2P^ = 9^. We can also verify easily that 2^ is a projection on a linear subspace if and only if (J^ - 2P) is a projection. If in particular 2P is the projection on Vi along V2, then (3 - ^) is the projection on V2 along Vi. In view of the preceding results, there is no ambiguity in simply saying a transformation 9^ is a projection (rather than 9^ is a projection on V\ along V2). We em-
phasize that if 2?^ is a projection, then
127 (2.104)
This is not necessarily the case for arbitrary linear transformations ST L(K V), for in general, 9l(2r) and J^CJ) need not be disjoint. Next, let ST G L(V; y). A linear subspace W of V is said to be invariant under the linear transformation ST if w G W implies that ^w G W. From this definition it follows trivially that (i) V is invariant under ST, (ii) {0} is invariant under ST, (iii) 2/1(9") is invariant under ST, and (iv) ^{'J) is invariant under ST. Next, let y be a linear space that is the direct sum of two linear subspaces W and U. If W and U are both invariant under a linear transformation ST, then ST is said to be reduced by W and U. It is readily verified, using definitions, that ST G L{V, V) is reduced by W and U if and only if S^ST = ST9^, where 9^ is the projection on W along U. Next we consider briefly the matrix representation of projections. To this end, let V be an ^-dimensional vector space and let 9^ G L( V, V). It is easily verified [using (2.104)] that if 9^ is a projection, then there exists a basis {v^ . . . , v'^} for V such that the matrix P of 9^ with respect to this basis is of the form (2.105)
1 0
0 1
0 0
>r
0 0
1
0
0
0
0
0
0 where r = dim 91(9^). We conclude with the following interesting result. Let V be a finite-dimensional vector space and let ^ G L{V, V). If W is a/^-dimensional invariant subspace of V and ifV = W®U, then there exists a basis for V such that the matrix A of ^^i with respect to this basis is of the form M2
A = 0
(2.106)
122
where An is a. p X p matrix and the remaining submatrices are of appropriate dimension. The canonical form (2.106) will be used in Chapter 3 in developing standard forms for uncontrollable and unobservable systems.
L. Some Canonical Forms of Matrices In this subsection we investigate under which conditions a linear transformation of a vector space into itself can be represented by (i) a diagonal matrix, (ii) a block
CHAPTER 2 :
Response of Linear Systems
128 Linear Systems
diagonal matrix, (iii) a triangular matrix, and (iv) a companion matrix. We will also investigate when a linear transformation cannot be represented by a diagonal matrix. Throughout this subsection, V denotes an fz-dimensional vector space over a field F. Diagonal form We begin with the following fundamental result. Let A i , . . . , A^ be the distinct eigenvalues of a linear transformation ^ G L(V, V) and let v^ 7^ 0 , . . . , v^ 7^ 0 be corresponding eigenvectors of si. Then it is easily shown that the set {v^ . . . , v^} is linearly independent. We note that if in particular ^ has n distinct eigenvalues, then the corresponding n eigenvectors span the linear space and, as such, form a basis forV. The above gives immediate rise to the next important result. Let si E L(V, V) and assume that the characteristic polynomial of si has n distinct roots, so that det (si - X3) = (Ai - A)(A2 - A)- • -(A^ - A),
(2.107)
where A i , . . . , A„ are distinct eigenvalues. Then there exists a basis {v^ . . . , v'^} of V such that v^ is an eigenvector corresponding to A/ for / = 1 , . . . , n. The matrix A of ^ with respect to the basis {v^ . . . , v"} is [Ai
0 A2
A = 0
= diag{Ki,...,\n\
(2.108)
Xn
In the same spirit as above, we can also easily establish the next result. Let si G L{V, V), and let A be the matrix of 6?^ with respect to the basis {v^ . . . , v"}. If the characteristic polynomial det {si — \3) = ao + ^ i A H h a^A" has n distinct roots, A i , . . . , \n, then A is similar to the matrix A of ^4 with respect to the basis {v^ . . . , v^}, where A is given in (2.108). In this case there exists a nonsingular matrix P such that A = P'^AP. The matrix P is the matrix of the basis {v^...,v"} with respect to the basis {v^ . . . , v"}, and P~^ is the matrix of the basis {v^ . . . , v^} with respect to the basis {v^ . . . , v"}. The matrix P can be constructed by letting its columns be eigenvectors of A corresponding to A i , . . . , A„, respectively; that is, P = [^1 ,^2^,^,^^«]^
(2.109)
where TT^ . . . , 77^ are eigenvectors of A corresponding to the eigenvalues A i , . . . , A„, respectively (verify this). The similarity transformation P given in (2.109) is called a modal matrix. If the conditions of the above result are satisfied and if, in particular, (2.108) holds, then we say that the matrix A has been diagonalized. As a specific case, let V be a two-dimensional vector space over the real numbers, let si G L{V, V), and let {v^ v^} be a basis for V. Suppose the matrix A of ^4 with respect to this basis is given by A =
\-2
.
4l
. . The characteristic polynomial
of ^ is p{X) = det(si - A^) = det{A - \I) = A^ + X-6 the eigenvalues of 62^ are A1 = 2 and A2 = - 3 .
= (\-
2)(A + 3), and
Let 77 = (r/i, 772)^ denote the coordinate representation of v G V witli respect to tlie basis {v , v }. To find eigenvectors corresponding to A/,/ = 1 , 2 , we solve the system of equations (A — A/) 77 = O. An easy computation yields 77^ = (1,1)^ and 77^ = (4, — 1) as eigenvectors corresponding to Ai and A2, respectively. The diagonal
0
matrix A given in (2.108) is A
The matrix P given in
0
(2.109) and its inverse P~^ are
[n\n" As expected, we have P
AP
4" -1
'1 1 "2 0
5
P-
1 _ ".2
.2
0" "Ai -3 = 0
.8" -.2
0' • By B y (2.81), the basis X2_
{ v \ v^} C y with respect to which A represents sz/ is given by v^ = Xj=i Pji ^^ = (r]i 772)^ is the coordinate representation of V with respect to { v \ v^}, then 77 P~^77 is the coordinate representation of V with respect to { v \ v^}. The vectors v \ v^ are of course eigenvectors of ^ corresponding to Ai and A2, respectively. When the algebraic multiplicity of one or more of the eigenvalues of a linear transformation is greater than one, then the linear transformation is said to have repeated eigenvalues. In this case it is not always possible to represent the linear transformation by a diagonal matrix. However, from the preceding results of this section it should be clear that a linear transformation with repeated eigenvalues can be represented by a diagonal matrix if the number of linearly independent eigenvectors corresponding to any eigenvalue is the same as the algebraic multiplicity of the eigenvalue. We consider two specific cases to shed additional light on this. First, we consider the matrix 1 0 0
3 4 3
-2 -2 -1
with characteristic equation det{A — XI) = (1—A)^(2 — A ) = 0 and eigenvalues Ai = 1 and A2 = 2. The algebraic multiplicity of Ai is two. Corresponding to Ai we can find two linearly independent eigenvectors (1,2,3)^ and (1,0,0)^, and corresponding to A2 we have an eigenvector (1,1,1)^. Letting P denote a modal matrix, we obtain 1 1 1 0 —1 1 1 2 1 2 0 1 , P-' = 3 0 1 0 3 -2
and
"1 A = P-^AP = 0 0
0 1 0
0" 0 2
In this example, dim ^ ^ = 2, which happens to be the same as the algebraic multiplicity of Ai. For this reason, we were able to diagonalize A.
129 CHAPTER 2 :
Response of Linear Systems
130 Linear Systems
As a second example, consider the matrix 2 0 0
1 2 0
-2 -1 1
with characteristic equation det(A — XI) = (1—A)(2 — A)^ = 0 and eigenvalues Ai = 1 and X2 = 2. The algebraic multiplicity of X2 is two and an eigenvector corresponding to Ai is (1,1,1)^. It is easily verified that any eigenvector corresponding to A2 must be of the form (vi, 0,0), Vi 7^ 0. We see that dim ^ ^ = 1, and thus, we are not able to determine a basis for the three-dimensional vector space V, which consists of eigenvectors. Consequently, we are unable to diagonalize A. Block diagonal form When a matrix cannot be diagonalized, we seek, for practical reasons, to represent a linear trasformation by a matrix that is as nearly diagonal as possible. The next result provides the basis of representing linear transformations by such matrices, called block diagonal matrices. In Subsection O of this section we will consider the "simplest" type of block diagonal matrix, called the Jordan canonical form. We let V be an n-dimensional vector space and we let ^ G L(y, V). IfV is the direct sum of p linear subspaces, Vi,..., Vp, which are invariant under ^ , then it can be readily shown that there exists a basis for V such that the matrix representation for ^ is in the block diagonal form given by
Ui I A =
0 Ai (2.110)
0 Moreover, A/ is a matrix representation of s^i, the restriction of ^ to Vi, / = 1,..., p. Also, p
det[A)=Wdet{A^).
(2.111)
i=\
From the above it is clear that to carry out a block diagonalization of a matrix A, we need to find an appropriate set of invariant subspaces of V, and furthermore, we need to find a simple matrix representation on each of these subspaces. As a specific case for the above, let V be an n-dimensional vector space. If ^ G L(y, y) has n distinct eigenvalues Ai,..., A^, and if we let ^ - = { v : ( ^ — Aj J^) v = 0}, 7 = 1,..., n, then jVj is an invariant linear subspace under ^ and V = ^ 0 • • • 0 JVn. For any v G JVJ, we have s^v = AyV, and hence, s^jV = A v for v G JVJ. A basis for ^j is any nonzero Vj G ^j. Thus, with respect to this basis, ^j is represented by the matrix Xj (in this case, simply a scalar). With respect to a basis of n linearly independent eigenvectors, { v \ . . . , v^}, ^ is represented by (2.108). Triangular form In addition to the diagonal form and the block diagonal form, there are many other useful forms of matrices to represent linear transformations on finite-dimensional vector spaces. One of these canonical forms involves triangular matrices with one of
the two forms given by
an 0
131
<3l2
ai3
.
ain
an
0
Cl22
(323
•
Clin
Cl2\
a22
or
0 0
0 0
0 0
. .
0 " 0
0 0
I
^n-l,n ^nn
<^n-l,l
Cln-1,2
Cln-1,3
^nl
^n2
UnS
' •'
0 0
(2.112) We call the matrix on the left an upper triangular matrix and the matrix on the right a lower triangular matrix. Now let V be an n-dimensional vector space over C, and let ^ G L(V,V)At can be shown that there exists a basis for V such that ^ is represented by an upper (or by a lower) triangular matrix with respect to that basis. We note that if A is in triangular form, then det(A - A/) = (an - A)(a22 - A)- • -(ann - A),
(2.113)
i.e., the diagonal elements of A are in this case the eigenvalues of A. Companion form We conclude this subsection by considering a canonical form for real square matrices that arises frequently in systems theory, called companion form. As a matter of fact, a given matrix A ^ R^^^ can be transformed via appropriate similarity transformations into four different forms of this type given by "0
/
X
X
"x '
/
x'
"0 x' / X
"x /
0 '
0 '
X
where the rows or columns denoted by (XX) are made up of the coefficients -ao, —ai,..., -an-i, determined by (-iydet(A - A/) = det(XI - A) = A'^ + an-iX""-^ + • • • + ao, where / G 7^("-i)x("-i), and where the 0 denotes an (n - 1)dimensional column or row vector. For example, the matrix on the left, which is perhaps the most commonly used companion form, and which henceforth we will identify as A^, is given by
0 0
1 0
0 1
..
0 0
0 -ao
0
0
..
1
— Ui
-(32
• ••
"fln-l
Ac =
In the following, we confine our discussion to this matrix. Similar treatments apply to the other three companion forms. Given A G R"-^"-, it can be shown that there exists a similarity transformation P that transforms A to the companion form Ac given above if and only if there exists a vector b EL R^ such that the matrix % = [b, Ab,..., A^~^b] is of full rank n. Furthermore, P is given by n-l
qA P =
[qAn-\
CHAPTER 2 :
Response of Linear Systems
^^^ Linear Systems
where q is the nth row of ^ - ^ and >. _ p - i / i p A matrix A for which such a vector Z? exists is called cyclic. It can be shown that A is cyclic if and only if the geometric multiplicity of each of its n eigenvalues is one, or equivalently, if and only if the n eigenvalues of A are exactly the roots of its minimal polynomial (refer to Subsections M and O, which follow). Finally, if A/ is an eigenvalue of Ac (or of A), then it can be verified that (1, A/,..., Ap^)^ is a corresponding eigenvector.
M. Minimal Polynomials One of our goals in this section is to develop the Jordan canonical form. To accomplish this we first need to introduce and study minimal polynomials (which will be accomplished in the present subsection) and nilpotent operators (to be considered in Subsection N). Throughout this subsection, V denotes an n-dimensional vector space. For purposes of motivation, consider the matrix
1 A = 0
0
3 -2 4 -2 3 -1
The characteristic polynomial of A is p(\) = (1 - A)^(2 - A), and we know from the Cayley-Hamilton Theorem that p(A) = O.
(2.114)
Now let us consider the polynomial m(A) = (1 - A)(2 - A) = 2 - 3A + A^. Then m(A) = 2/ - 3A + A^ - O.
(2.115)
Thus, matrix A satisfies (2.115), which is of lower degree than (2.114), the characteristic equation of A. More generally, it can be shown that for annX n matrix A there exists a unique polynomial m(A) such that (i) m(A) = O, (ii) m(A) is monic (i.e., if m is an nthdegree polynomial in A, then the coefficient of A" is unity), and (iii) if m\X) is any other polynomial such that m\A) = O, then the degree of m(A) is less or equal to the degree of m'(A) [i.e., m(A) is of the lowest degree such that m(A) = O]. The polynomial m(A) is called the minimal polynomial of A. In the remainder of this section we let p(X) denote the characteristic polynomial of an n X n matrix A and we let m(A) denote the minimal polynomial of A. In what follows, we develop an explicit form for the minimal polynomial of A that makes it possible to determine it systematically. Let /(A) be any polynomial such that /(A) = O (e.g., the characteristic polynomial). Then it is easily shown that m(A) divides /(A) [i.e., there is a polynomial ^(A) such that /(A) = ^(A)m(A)]. In particular, the minimal polynomial of A, m(A), divides the characteristic polynomial of A, p(X). Also, it can be shown that p(X) divides [m(A)]^
Next, let p(X) be given by
133
p(X) = (Ai - ArHA2 - A r • • -(A^ - A r ^
(2.116)
where m i , . . . , m^^ are the algebraic multiplicities of the distinct eigenvalues A i , . . . , \p of A, respectively. It can be shown that m(A) = (A - AO^KA - A2)^^---(A - A . r ^
(2.117)
where 1 < /i/ < mt, i = \,..., p. The only unknowns left to determine the minimal polynomial of A are )Lti,..., fjip in (2.117). These can be determined in several ways. The next result is a direct consequence of (2.86). Let A be similar to A and let m(A) be the minimal polynomial of A. Then m(A) = m(A). This result, in turn, justifies the following definition. Let d^ G L{V, V). The minimal polynomial of si is the minimal polynomial of any matrix A that represents M. To develop the Jordan canonical form (for linear transformations with repeated eigenvalues), we need to consider several additional preliminary results that are important in their own right. Let ^ G L(V; V) and let /(A) be any polynomial in A. Let M'f = {v : f(A)v = 0}. It can be shown that Xf is an invariant linear subspace of V under ^ . In particular let A i , . . . , Ap be the distinct eigenvalues of ^4 G L(V, V) and for j = 1 , . . . , p, and for any positive integer q, let Xj = {v : ( ^ - Xj^)^v = 0}.
(2.118)
In view of the above result, it follows that Xj is an invariant linear subspace of V under ^ . Next, let ^ G L(V, V), let Vi and V2 be Hnear subspaces of V such that V = Vi 0 V2, and let ^ 1 be the restriction of 6?^ to Vi. Let /(A) be any polynomial in A. It can be shown that if si is reduced by Vi and V2 then, for all v^ G V\, f(sii)v^ =
f(d)vK Next, let y be a vector space over ^ and let M G L(V,V). Let m(A) be the minimal polynomial of ^ as given in (2.117). Let ^(A) = (A - Ai)^i, let /z(A) = (A - A2)^2.. .(A - Ap)^/' if/? > 2, and let /z(A) = 1 if p = 1. Let 6^1 be the restriction of si to jNff S i.e., div = siv for all v G >f/'\ Let J/t = {v G V : h(si)v = 0}. Using the preceding results, we can show that (i) V = X{^^ ®M, and (ii) (A - AO^^ is the minimal polynomial of ^ 1 . These make it possible to prove the next result, which is called the primary decomposition theorem for linear transformations and which we state next. Let V be an n-dimensional vector space over C, let A i , . . . , A^ be the distinct eigenvalues ofsi^ L{V, V), let the characteristic and minimal polynomials of ^ be k^p
and
m(A) = (A - AO^^ • • -(A - A.)^^
respectively. Let Vi = {v :
A^-i)A^'v = 0},
i =
l...,p.
(2.119)
Then (i) Vi, i = 1,...,/? are invariant linear subspaces of V under si, (ii) V = Vi © • • • 0 Vp, (iii) (A - A/)^' is the minimal polynomial of sit, where ^i is the restriction of si to Vi, and (iv) dim Vi = mi, i = \,..., p.
CHAPTER 2 :
Response of Linear Systems
134 Linear Systems
The above result shows that we can always present ^ G L(y, V) by a matrix in block diagonal form, where the number of diagonal blocks is equal to the number of distinct eigenvalues of ^ . We will next consider a convenient representation for each of the diagonal submatrices A/. It may turn out that one or more of the submatrices A/ will be diagonal. The next result tells us specifically when ^ G L(y, V) is representable by a diagonal matrix. Let V be an n-dimensional vector space over C and let ^ G L(V,V). Let Ai,...,Ap,p < n, be the distinct eigenvalues of ^ . Then there exists a basis for V such that the matrix A of ^ with respect to this basis is diagonal if and only if the minimal polynomial for ^ is of the form m(A) = ( A - A i ) ( A - A 2 ) - - - ( A - A p ) .
(2.120)
N. Nilpotent Operators Let us now proceed by considering a representation for each of the ^ - G L(Vi-, Vi) in the primary decomposition theorem presented in Subsection M so that the block diagonal matrix representation of ^ G L(y, V) is as simple as possible. To accomplish this, we need to define and examine nilpotent operators. Let ^ G L(y, V). Then ^ is said to be nilpotent if there exists an integer q>Q such that ^^ = ^ . A nilpotent operator is said to be of index q if ^^ = ^ but Recall that the primary decomposition theorem enables us to write V = Vi 0 • • • 0 Vp. Furthermore, the linear transformation (^- — Xi/) is nilpotent on Vi. If we let JVi = £^i — Xi/, then ^- = A / ^ + JVi. Now A / ^ is clearly represented by a diagonal matrix. However, the transformation jVi forces the matrix representation of ^ - to be in general nondiagonal. Therefore, the next task is to seek a simple representation of the nilpotent operator jVi. In the next few results, which are concerned with properties of nilpotent operators, we drop the subscript / for convenience. Let JV G L(W, W ) , where W is an m-dimensional vector space. It can be shown that if ^ is a nilpotent linear transformation of index q and if w G W is such that JV^~^W 7^ 0, then the vectors w, JVw^..., JV^~^w in W are linearly independent. Next, we examine the matrix representation of nilpotent transformations. Let W be a ^-dimensional vector space and let JV G L(W,W) be nilpotent of index q. Let w^ G W be such that JV^~^w^ ^ 0. It can be shown that the matrix N of ^ with respect to the basis {^^ •VO, A'l-^w^,. .,Mp}^m^^ is [J^i0 0
1 0
0 1
0 0
... ...
0 0
0 0
0 0
0 0
0 0
0 0
... ...
0 0
1 0
(2.121)
N-
The above result characterizes the matrix representation of a nilpotent linear transformation of index ^ on a ^-dimensional vector space. The next task is to determine the representation of a nilpotent operator of index 7 on a vector space of dimension m, where y < m. It is easily shown that we can dismiss the case y > m, i.e., if ^ G L(W, W) is nilpotent of index 7, where dim W = m, then y<m.
Now let W be an m-dimensional vector space, let J{ G L(W,W), let y be any positive integer, and let
W2 = {w\ M^w = 0}, dim W2 = h (2.122) Wy = {w: X^w = 0}, dim W^ = L. Also, for any / such that 1 < / < 7, let {w^,..., w^} be a basis for W such that {w^,..., w^'} is a basis for Wi. It can be shown that (i) Wi C W2 C --- C Wy and (ii) {w^,..., w^'-i, Xw^^^^,..., Xw^^+^} is a linearly independent set of vectors in Wi. The next result, which is the principal result of this subsection, is a consequence of the above results. Let W be an m-dimensional vector space over C and let X ^ L(W, W) be nilpotent of index 7. Let Wi = {w : J^w = 0},.. .,Wy = {w : Ji^w = 0}, and let // = dim Wi, i = 1 , . . . , 7. Then there exists a basis in W such that the matrix A^ of >r is of block diagonal form. 0 (2.123)
A^ -
Nr
where
Ni
0 0
1 0
0 1
0 0
. . . .
0 0
0 0
0 0
0 0
0 0
0 0
. . . .
0 0
1 0
(2.124)
=
/ = 1 , . . . , r, where r == /i, Ni is a ki X ki matrix, 1 < fc^ < 7, and ki is determined in the following manner: there are ly - ly_i 2li - li+i - li-i 2/1-/2
7 X 7 matrices, i X / matrices, 1 x 1 matrices.
/ = 2 , . . . , 7 - \,
(2.125)
The basis for W consists of strings of vectors of the form M^^ ^w^,..., w^ J^^^ iy^;2^ .. .,w' \ . . .,K^'
^W^
CHAPTER 2 :
Response of Linear Systems
W^ = {w: J^w = 0}, dim Wi = h
A^i
135
...,W\
O. The Jordan Canonical Form The results of the preceding three subsections can be used to prove the next result, which yields the Jordan canonical form of matrices. Let Y be an /i-dimensional vector space over C and let ^ G L(V; V). Let the characteristic polynomial of ^4 be p{K) = (Ai - X)^^ • • -(A^ - A)^^ and let the minimal polynomial of d be m(A) == (A - Ai)^i • • -(A - A^,)^^, where A i , . . . , Ap are the distinct eigenvalues of 6^. Let Vi = {v G V : (si-Ai3)^^v = 0}. Then (i) Vu...,Vp are invariant subspaces of y under 5i; (ii) y == y i © - - - © ^ ^ ; (iii)dim V/ = m/, / = \,..., p\ and (iv) there exists a basis for V such that the matrix Aof d^ with respect
136
to this basis is of the form
Linear Systems A =
"Ai 0 0
0 • • A2 • • 0
0 0
(2.126)
• • Ap
where A,- is an m, X m,- matrix of the form Ai - A// + Ni
(2.127)
and where Ni is the matrix of the nilpotent operator {sit - A/^) of index [xi on V/ given by (2.123) and (2.124). Parts (i) to (iii) of the above result are restatements of the primary decomposition theorem. From this theorem we also know that (A - A/)^' is the minimal polynomial of sii, the restriction of d^ to Vi. Hence, if we let Mi = sii-Xi3, then J^T/ is a nilpotent operator of index />t/ on Vi. We are thus able to represent Xi as shown in (2.124). A little extra work shows that the representation of ^ G L(V; V) by a matrix A of the form given in (2.126) and (2.127) is unique except for the order in which the block diagonals Ai,., ,Ap appear in A. The matrix A of ^ G L(K V) given by (2.126) and (2.127) is called the Jordan canonical form ofd. An example We conclude this section by considering a specific case. We let V = R^, we let {e^,..., ^^} be the natural basis for V, and we let, L(V, V) be represented by the matrix
A -
1 0 2 2 0 0 1
0 1 1 0 0 0 -1
-1 0 2 -1 0 0 0
1 0 -1 2 0 0 1
1 0 -1 1 1 0 2
3 0 -6 3 0 1 4
0 0 0 0 0 0 1
with respect to {e^,..., e^}. We wish to determine the matrix A that represents si in the Jordan canonical form. We first determine that the characteristic polynomial of si is det (A - A/) = det(si - \3) = (1 - A)^. This indicates that Ai = 1 is the only distinct eigenvalue of d, having algebraic multiplicity m\ = 7. To find the minimal polynomial of si, we let >r = ^ - Ai^, where 3 is the identity operator in L{V,V). The representation of Ji with respect to the natural basis is
N = A-I
-2 0 2 = -2 0 0 -1
0 0 1 0 0 0 -1
-1 0 1 -1 0 0 0
1 0 -1 1 0 0 1
1 0 -1 1 0 0 2
3 0 -6 3 0 0 4
0 0 0 0 0 0 0
The minimal polynomial will be of the form m(A) = (X-iyK We need to determine the smallest vi such that m(A — A/) = m(N) = O. We first obtain
'2
_
"0 - 1 0 0 0 1 0 -1 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
3 0 -3 3 0 0 0
0 0 0 0 0 0 0
Next, we obtain N^ = 0, and therefore, z^i = 3 and Jvf is a nilpotent operator of index 3. We see that V = J{f [refer to (2.118) for the notation >ff]. We will now apply the results of Subsection F to obtain a representation for M in this space. We let Wi = {v : Xv = 0}, W2 = {v : J^^v = 0}, and W3 = {v : J{\ = 0}, and we observe that A^ has three linearly independent rows. This means that the rank of jf is 3, and therefore, dim(Wi) = h = 4. Similarly, the rank of X^ is 1 (since N^ has one linearly independent row), and so dim (^2) = h = 6. Clearly, dim(W3) = h = 1. We conclude that }( will have a representation N of the form (2.123) with r = A. Each of the Nt will be of the form (2,124). There will be h-h = 1 (3 X 3) matrix, Ik - h - h = 1 (2 X 2) matrix, and 2/i - fc - 2 (1 X 1) matrices [see (2.125)]. Hence, there is a basis for V such that >f may be represented by the matrix 0 0 0 7V=|0 0 0 0
1 0 0 0 0 0 0
0
0 1 0 0 0 0 0 0 0 0 0 0 0
0 0 0
0 0 0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0
The corresponding basis will consist of strings of vectors of the form X'^v^, Xv^, v^ >fv^, v^, v^, v^. We will represent the vectors v^v^, v-^, and v^ by 17^17^,17^, and iq^, their coordinate representations, respectively, with respect to the natural basis {e^, e^, e^, e^, e^, e^, e^} in V. We begin by choosing v^ G W3 such that v^ ^ W2, i.e., we determine a iq^ such that N^rj^ = 0 but N^r]^ 9^ 0. The vector TJ^ = (0, 1, 0, 0, 0, 0, 0)^ will do. We see that Nr]^ = (0, 0, 1, 0, 0, 0, - 1 ) ^ and A^^^^ = ( - 1 , 0, 1, - 1 , 0, 0, 0)^. Hence, Jiv^ E W2 but Xv^ ^ Wi and X^v^ G Wj. We see there will be only one string of length 3, and therefore we choose next v^ E W2 such that v^ ^W\. Also, the pair {Jfv^ v^} must be linearly independent. The vector T/2 - (1, 0, 0, 0, 0, 0, 0)^ will do. Next, we have Nr]^ - ( - 2 , 0, 2, - 2 , 0, 0, - 1 ) ^ , and therefore, Xv'^ E Wi. We complete the basis for V by selecting two more vectors, v^, v"^ E Wi, such that {Ji^v^, M'v^, v^, v^} are linearly independent. The vectors 7]^ = (0, 0, - 1 , - 2 , 1, 0, 0)^ and ry^ = (1, 3, 1, 0, 0, 1, 0)^ will do. It now follows that the matrix P = [N^rj^, Nrj^, 17^ Nrj'^, 17^, 17^, r]"^] is the matrix of the new basis with respect to the natural basis. The reader can readily verify
137 CHAPTER 2 :
Response of Linear Systems
138
thatA^ = P~^NP,whQYQ
Linear Systems
P =
1 0 0 0 1 1 1 0 0 0 0 0 0 -1
0 -2 1 0 2 0 2 0 0 0 0 0 0 -1
1 0 1 0 0 3 0 -1 1 0 - 2 0 ,p-' 1 0 0 0 0 1 0 0 0
0 0 0 = 0 1 0 0
0 2 1 4 1 1 0 3 1 0 0 0 0 -1 -1 -3 0 0 -1 -2 0 1 0 0 0 0 0 0
-2 2 -1 0 -3 0 1 -1 -1 0 0 0 1 0
Finally, the Jordan canonical form for A is now given by A = A^ + / (recalling that the matrix representation for ^ is the same for any basis in V). Therefore, we have 1 0 0 A = 0 0 0 0
1 1 0 0 0 0 0
0:0 0 0 1 i0 0 0 1 1 0 0 0 o! 1 1 0 O'O 1 0 0 0 0 1 0 0 0 0
0 0 0 0 0 0 1
It is easily verified that A = P ^AP. In general it is more convenient as a check to show that PA = AP.
2.3 LINEAR HOMOGENEOUS AND NONHOMOGENEOUS EQUATIONS In this section we consider systems of linear homogeneous ordinary differential equations X = A(t)x
(LH)
and linear nonhomogeneous ordinary differential equations X = A(t)x + ^(0.
(LN)
In Theorem 12.1 of Chapter 1 it was shown that these systems of equations, subject to initial conditions x(to) = XQ, possess unique solutions for every (to, XQ) E D, where D = {(t, x):t GJ = (a,b),xE R""} and where it is assumed that A G C(/, i?"^'") and g E C(J, R^), These solutions exist over the entire interval / = (a, b) and they depend continuously on the initial conditions. Typically, we will assume that / = (-00, 00). We note that (/)(0 = 0, for all r E / , is a solution of (LH), with (l)(to) = 0. We call this the trivial solution. As in Chapter 1 (refer to Section 1.13), we recall that the preceding statements are also true when A(t) and g(t) are piecewise continuous on/. In the sequel, we sometimes will encounter the case where A(t) = A is in Jordan canonical form that may have entries in the complex plane C. For this reason, we will allow D = {(t, x):t E J = (a,b\xE /?"(or x E C")} and A E C(J, JR"^^'^) [or A E C(J, C^^^)], as needed. For the case of real vectors, the field of scalars for
the X-space will be the field of real numbers {F = R), while for the case of complex vectors, the field of scalars for the x-space will be the field of complex numbers ( F = C). For the latter case, the theory concerning the existence and uniqueness of solutions for (LH), as presented in Chapter 1, carries over and can be modified in the obvious way. A. T h e F u n d a m e n t a l M a t r i x Solution space We will require the following result. THEOREM 3.1. The set of solutions of (LH) on the interval / forms an ^-dimensional vector space. Proof, Let V denote the set of all solutions of (LH) on / . Let a i , a 2 G F and let 01,02 G V. Then a i 0 i + 0(202 ^ V since {d/dt)[ai(l)i + 0(202] = a\{d/dt)(\)\{t) + a2{d/dt)(^2{t) = aiA(r)0i(r) + a2A(r)02(O = A(r)[ai0i(r) + 0(202(0] for ^^ ^ ^ JThis shows that V is a linear subspace of/?". Hence, V is a vector space. To complete the proof of the theorem, we must show that V is of dimension n. To accomplish this, we must find n linearly independent solutions 0 i , . . . , 0„ that span V. To this end, we choose a set of n linearly independent vectors XQ,...,XQ in the ^-dimensional x-space (i.e., in R"^ or C"). By the existence results in Chapter 1, if ^o ^J^ then there exist n solutions 0 i , . . . , 0„ of (LH) such that 0i (^o) = -^O' • • •' ^«(^o) = -^0- ^ ^ first show that these solutions are linearly independent. If on the contrary, these solutions are linearly dependent, there exist scalars a i , . . . , a„ G F, not all zero, such that E L i ^?0KO = 0 for all t G / . This implies in particular that ^4=1 oci^i{to) = E L i ^i^o ^ 0. But this contradicts the assumption that {XQ, ... ,XQ} is a linearly independent set. Therefore, the solutions 0 i , . . . , 0„ are linearly independent. To conclude the proof, we must show that the solutions 0 i , . . . , 0„ span V. Let 0 be any solution of (LH) on the interval / such that 0(^o) = -^o- Then there exist unique scalars a i , . . . , a„ G F such that x o ••
Z^^i^Qy
since, by assumption, the vectors XQ, ... ,XQ form a basis for the x-space. Now n
is a solution of (LH) on / such that \j/{^o) = ^o- ^^t by the uniqueness results of Chapter 1 we have that „
Since 0 was chosen arbitrarily, it follows that the solutions 0 i , . . . , 0„ span V.
•
EXAMPLE 3.1. Let A (t) for (LH) be given by
Ait):
-1 0
e^n -1
(3.122)
It is easily verified by direct substitution that 0i(O = (e \0) and (1)2 = {\e\e ^) are solutions oi {LH) onJ= (—00^00) for the present case. Furthermore, it is easily
139 CHAPTER 2: Response of Linear Systems
140 Linear Systems
shown that (pi and (p2 [defined on / = (—^,^)] are hnearly independent (using, e.g., the method in Subsection 2.2B). In view of Theorem 3.1, the solutions of (LH) with A{t) specified by (3.1) form a two-dimensional vector space and {0i, ^2} is a basis for this solution space. Since {0i,02} spans this vector space, all solutions of (LH) with A{t) specified by (3.1) are of the form ^(t) = ai^i{t) + ^202(0 = {oc\e~^ + 0^2 '].\J e^, 0^2^ 0 , where a i , 0^2 G R.
F u n d a m e n t a l m a t r i x a n d properties Theorem 3.1 enables us to make the following definition. DEFINITION 3.1. A set of n linearly independent solutions of {LH) on / , { 0 i , . . . , 0„}, is called di fundamental set of solutions of {LH), and ihQnxn matrix O=[0i,02,...,0n] "011 021
012 022
••• •••
01n 02n
Pnl (t>n2 is called a. fundamental matrix of {LH). We note that there are infinitely many different fundamental sets of solutions of {LH) and, hence, infinitely many different fundamental matrices for {LH). We now study some of the basic properties of fundamental matrix. In the next result, X = [xij] denotes an n x n matrix, and the derivative ofX with respect to t is defined as X = [xij]. Let A{t) be the nx n matrix given in {LH). We call the system of n^ equations X=A{t)X (3.2) a matrix differential
equation.
THEOREM 3.2. A fundamental matrix O of {LH) satisfies the matrix equation (3.2) on the interval / . Proof, WehaveO=[0i,02,...,0n] = [A(O0i,A(O02,...,A(O0n]=A(O[0i,02,...,0n] = A{t)^. • The next result is cailcd Abel's
formula.
THEOREM 3.3. If O is a solution of the matrix equation (3.2) on an interval / and T is any point of / , then det 0{t) = det 0 ( T ) exp
/
trA{s)ds
for every t e J. [tr A{s) = tr [aij{s)] denotes the trace of A{s), i.e., tr A{s) I"=i«;7(*)-]
141
Proof, Let ^ = [(j^ijl Then ^tj = Zl=i atkit^kj^ Now
dt
[det<^(t)] =
CHAPTER 2 :
>ii
12
• ..
021
022
. ..
>nl
0n2
•
f ••• +
01« 02n
+
.. 0«n
011 021
012 022
0nl
0«2 •
011 021
012 022
..• • ••
0nl
0n2
• •
01n 02n
Response of Linear Systems
. - . 01« • • • 02n
0nn
021
022
0nl
0-•n2
T.l^iaik{t)(j>kn{t) 02n (t>nn
011 Y.l=iCl2k{t)4>k\
012 2 ^ = 1 a2k(t)
031
032
01« Y.l=iCl2k{t)4>kn 4>'in
0nl
0ril
4>n
•
012 022
011 021
01n 02n
+ 0n-l,l
0«-l,2
\Y.l=\Clnk(t)4>kl
...
Y.l=\Clnk{t)(f)k2
••.
0n-l,n YJ[=iClnk{t)(j>kn
Thefirstterm in the above sum of determinants is unchanged if we subtract from the first row the quantity {an times the second row) + (^13 times the third row) + \-{ain times the nth row). This yields 011 021
012 022
... ...
01n 02n
(2ii(O011 021
ai l(O012
022
. .
a\ l(O01«
02n
(knl
0n2
(f>nn
zzz
0nl
0n2
.
.. 0«n =
2n(t)det<^(t).
i
Repeating the above procedure for the remaining terms in the above sum of determinants, we have (d/dt)[det ^(t)] = an(t)det ^(t) + a22(t)det ^(t) + • • • + UnniOdet ^ ( 0 = [tr A(t)] det ^{t). This imphes that det (0 = det ^(r) exp [{/ tr A{s) ds]. m Since in Theorem 3.3 r is arbitrary, it follows that either det 0 ( 0 7^ 0 for all t E J or det 0 ( 0 = 0 for each t ^ J. The next result provides a test on whether an nX n matrix 0 ( 0 is a fundamental matrix of {LH), THEOREM 3.4. A solution O of the matrix equation (3.2) is a fundamental matrix of {LH) if and only if its determinant is nonzero for alH E / .
142 Linear Systems
Proof, If = [(/>i, (/)2,..., (/>n] is a fundamental matrix for (L//), then the columns of $ , (/)i,..., (/>„, form a hnearly independent set. Now let (/> be a nontrivial solution of {LH). Then by Theorem 3.1 there exist unique scalars a i , . . . , a„ E F , not all zero, such that (f) = 2 ; = ! oij4'j = ^^» where a^ = (ai,..., an). Let t = T G J. Then ^ ( T ) = (T)(2, which is a system of n linear algebraic equations. By construction, this system of equations has a unique solution for any choice of (^(r). Therefore, det 0 ( T ) T^ 0. It now follows from Theorem 3.3 that det ^(t) ^ 0 for any t G / . Conversely, let ^ be a solution of (3.2) and assume that det ^{f) T^ 0 for all ^ G / . Then the columns of are linearly independent for all ? G / . Hence, ^ is a fundamental matrix of {LH). • It is emphasized that a matrix may have identically zero determinant over some interval, even though its columns are linearly independent. For example, the columns of the matrix
'1 0(0 = 0 0
t 1 0
f t 0
are linearly independent, yet det 4>(0 = 0 for all t G (-oo, oo). In accordance with Theorem 3.4, the above matrix cannot be a fundamental solution of the matrix equation (3.2) for any continuous matrix A{f). EXAMPLE 3.2. Using the linearly independent solutions
0(0
2"
0
e~
t G ( - 0 0 , 00),
as a fundamental matrix of {LH) with A{t) given by (3.1). It is easily verified that the matrix O satisfies the matrix equation O = AO. Furthermore, det 0 ( 0 = e~^^ ¥^ 0,t G (-00,00), and det 0 ( 0 = det 0 ( T ) e x p [ | / r r A{s)ds] = ^ ~ ^ ^ e x p [ | / - 2 J^] = ^-2r^-2(r-T) ^ ^-2t^ f ^ ^_Qo^ 00^^ ^g expected. • THEOREMS.5. If O is a fundamental matrix of {LH) and if C is any nonsingular constant nXn matrix, then OC is also a fundamental matrix of {LH). Moreover, if ^ is any other fundamental matrix of {LH), then there exists a constant nX n nonsingular matrix P such that ^ = OP. Proof. For the matrix OC we have {d/dt){^C) = OC = [A(0O]C = A(0(OC), and therefore, OC is a solution of the matrix equation (3.2). Furthermore, since det 0 ( 0 ^ 0 for ? G / and det C 7^ 0, it follows that det [O(0C] = [det ^{t)]{det C) 7^ 0, ^ G 7. By Theorem 3.4, OC is a fundamental matrix. Next, let ' ^ be any other fundamental matrix of {LH) and consider the product O'HO"^. [Notice that since det^{t) 7^ 0 for all t G 7, then O"H0 exists for all t G J.] Also, consider 0 0 ~ ^ = /, where / denotes the ^ X n identity matrix. Differentiating both sides, we obtain [(
(33)
has two linearly independent solutions given by 4>i(t) = {e^\ e^^Y, <^2(0 ^ (^^ 2^0^. 143 and therefore, the matrix CHAPTER 2 : Response of (3.4) ^(0 Linear Systems 2e' is a fundamental matrix of (3.3). Using Theorem 3.5 we can find the particular fundamental matrix "^ of (3.3) that satisfies the initial condition '^(0) = / by using 0(0 given in (3.4). We have '^(0) = / = O(0)C or C = 0-1(0), and therefore, C =
and
II
I I
^ ( 0 = OC =
{2e^' -e') i-e^' -he') {2e^' - 2e') {-e^' + 2e')
B. The State Transition Matrix In Chapter 1 we used the Method of Successive Approximations (Theorem 10.9) to prove that for every {t^, XQ) G J X R^, X = A(t)x
(LH)
possesses a unique solution of the form (/)(/, to, Xo) = ^(t,
to)Xo,
such that (l)(to, to, xo) = xo, which exists for all t G J, where <E>(r, ^o) is the state transition matrix (see Section 1.13). We derived an expression for ^(t, to) in series form, called the Peano-Baker series [see Eq. (13.3) of Chapter 1], and we showed that 0(r, ^o) is the unique solution of the matrix differential equation ^-^(t, to) = A(t)^(t, to\ dt
where
^{to, to) = I
(3.5)
(3.6)
for all t G /
We provide an alternative formulation of state transition matrix and we study some of the properties of such matrices. In the following definition, we use the natural basis {ei, e2,..., Cn) that was defined in Section 2.2. DEFINITIONS.2. A fundamental matrix O of (LH) whose columns are determined by the linearly independent solutions (^i,..., (/>„ with ^ l ( ^ ) = ei, . ..,(f)n{tQ) = en.
to E / ,
is called the state transition matrix O for (LH). Equivalently, if ^ is any fundamental matrix of (LH), then the matrix O determined by (t, to) = ^ ( 0 ^ " k ^ )
for all t, to G /,
is said to be the state transition matrix of(LH). We note that the state transition matrix of (LH) is uniquely determined by the matrix A(t) and is independent of the particular choice of the fundamental
•
144 Linear Systems
matrix. To show this, let ^ i and ^ 2 be two different fundamental matrices of (LH). Then by Theorem 3.5 there exists a constant n X n nonsingular matrix P such that ^ 2 = ^iP. Now by the definition of state transition matrix, we h a v e 3 ) ( U o ) = ^ 2 ( 0 [ ^ 2 ( ^ ) ] - ' = %(t)PP-'[%(to)]-' = ^ i ( 0 [ ^ i a o ) ] " ^ This shows that ^(t, to) is independent of the fundamental matrix chosen. EXAMPLE 3.4. In Examples 3.1 and 3.2 we showed that for (LH) with A(t) given by (3.1),
^(0 =
2^
-X
tG(-
0
is a fundamental matrix. For this case, the state transition matrix ^(t, to) is computed as ^(t,to) =
^(t)^-\to)
0
0 -t+3to
The reader should verify that this matrix satisfies Eqs. (3.5) and (3.6). Properties of the state transition matrix In the following, we summarize some of the properties of state transition matrix. THEOREM 3.6. Let ^o ^ J, let (l>(to) = XQ, and let 0(^, ^o) denote the state transition matrix for (LH) for all / E J. Then the following statements are true: (i) 0(f, ^0) is the unique solution of the matrix equation (dldt)^(t, to) = A(t)^(t, to) with 0 ( ^ , to) = I,the nX n identity matrix, (ii) (r, ^0) is nonsingular for all ^ E / . (iii) For any t,a,TE / , we have (^(t, r) = ^(t, cr)^(cr, r) (semigroup property), (iv) [(t, to)V^ = ^-\t, to) = <^(to, t) for all t, to E / . (v) The unique solution (/>(r, to, xo) of (LH), with (/)(^, ^o, xo) Xo specified, is given by 4>(t, to, Xo) = ^(t, to)xo
for all t E /.
(3.7)
Proof (i) For any fundamental matrix of (LH), say, ^, we have, by definition, ^(t, to) = ^ ( 0 ^ ~ H ^ ) , independent of the choice of "¥. Therefore, d^(t,to)ldt = ^(t)'^-\to) = A(t)^(t)^-\to) = A(t)^(t,to)' Furthermore, ^(toJo) = ^(to)^-\to) = L (ii) For any fundamental matrix of (LH) we have that det "^(t) 7^ 0 for all ^ E 7. Therefore, det ^(t, to) = det [^(0^~H^o)] = det 'i^(t)det'i^~\to) 7^ 0 for all t, to E / . (iii) For any fundamental matrix "^ of (LH) and for the state transition matrix O of (LH), wehave^(r,T) = - ^ ( 0 ^ " ^ ^ ) = 'i^(t)^-\a)'¥(a)^~\T) = ^(t,a)(a,r) for any t,a,TE. J. (iv) Let '^ be any fundamental matrix of (LH) and let O be the state transition matrix of (LH). Then [^(tJo)]'^ = WO^(^o)"M"^ = ^ ( r o ) ^ " k O = 0 ( ^ , 0 for any t, to E 7. (v) By the results established in Chapter 1, we know that for every (^0. ^0) ^ D, (LH) has a unique solution 0 ( 0 for all r E 7 with )(?o) = XQ. To verify (3.7), we note that 4)(t) = [d^(t, to)/dt]xo = A(t)<^(t, to)xo = A(t)(l)(t). m
EXAMPLE 3.5. Let x(^o) = (ai, a2)^. In view of Example 3.4, we have for (L//) given in Example 3.1, (/)(r, to, Xo) =
0
and therefore, (3.8) (3.9)
We can verify the above example by simple integration. In doing so, we first obtain (^2(^> ^> -^o) by integrating both sides of i:2 "= ~-^2 and by using the initial condition X2(to) = CLI- Next, we obtain (piit, to, XQ) by integrating both sides of i i = -x\ + e^^cl)2(t, to, XQ) and using the initial condition (x\(to), ^2(^0))^ = (ai, ^2)^. Note that this procedure for solving an initial-value problem determined by (LH) is valid for any triangular matrix A(t). In Chapter 1 we pointed out that the state transition matrix ^(t, to) maps the solution (state) of (LH) at time to to the solution (state) of (LH) at time t. Since there is no restriction on t relative to to (i.e., we may have t < to,t = to, or t > to), we can "move forward or backward" in time. Indeed, given the solution (state) of (LH) at time t, we can solve the solution (state) of (LH) at time to- Thus, x(to) = xo = [^(t, to)V^(l>(t, to, Xo) = 0(^0, t)(l)(t, to, Xo). This "reversibility in time" is possible because ^~^(t, to) always exists. [In the case of discrete-time systems described by difference equations, this reversibility in time does in general not exist (refer to Section 2.7).]
C. Nonhomogeneous Equations In Section 1.13, we proved the following result [refer to Eqs. (13.8) to (13.10)]. THEOREM 3.7. Let to E 7, let (to, Xo) E D, and let ^(t, to) denote the state transition matrix for (LH) for all r E / . Then the unique solution (/)(r, to, xo) of (LN) satisfying >(^, to, Xo) = Xo is given by ^(^, ^0, ^o) = ^(t, to)xo +
^(t, v)8(v)dV'
(3.10)
As pointed out in Section 1.13, when xo = 0, (3.10) reduces to ct>(t,to,0) ^ cl>p(t) = f ^(t,s)g(s)ds.
(3.11)
Jto
and when xo # 0, but g(t) = 0, (3.10) reduces to ct)(t, to, Xo) = (l>h(t) = ^(t, to)xo.
CHAPTER 2 :
Response of Linear Systems
^-(^-^o)
(t>2(t, to, Xo) = e-^'-'^^a2.
145
(3.12)
and the solution of (LN) may be viewed as consisting of a component that is due to the initial data xo and another component that is due to the forcing term g(t). We recall that cffp is called a particular solution of the nonhomogeneous system (LN), while (f)h is called the homogeneous solution. There are of course other methods of solving (LN). For example, in the common approach to solving linear differential equations, all solutions (^(0 of i - A(t)x =
146 Linear Systems
g{t) are assumed to be of the form 0(f) = ^^{t) + ^p{t), where 0/^(f) is the solution of the homogeneous equations x — A{t)x = 0 and ^p{t) is a particular solution. It is not difficult to see that 0/^(f) = 0(f, to)oc, where a eR^ is to be determined [compare to (3.12)]. Therefore, (p{t) = <^{t,to)a-\- (pp{t). The vector a is determined using the initial conditions on the solution (p{t) of x — A{t)x = g{t), namely, (p{to) = XQ. Substituting (j){to) = O(fo,^o)oj + 0p(fo), we obtain a = (j){to) - 0p(fo) = -^o - (t>p{to)' The solution to (LN) with 0(fo) = -^o is therefore given by 0(f) = 0/^(f) + ^p{t) = O(f,fo)[0(^o) - 0p(^o)] + 0p(O' or 0(O=O(f,fo)^o + [0p(O-^(^^o)0p(^o)].
(3.13)
As expected, the "zero-state response of {LN)'' obtained by letting XQ = 0 in (3.10) and given in (3.11), is a particular solution ^p{t) ofx—A{t)x = g{t) with the property that 0p(fo) = 0- Furthermore, the "zero-input response of(LN)'' obtained by letting g(t) = 0 in (3.10), is the solution 0/^(f) ofx — A{t)x = 0. It follows that the variation of constants formula (3.10) is a special form of expression (3.13), corresponding to the chosen particular solution (pp{t)', for a different choice of (pp{t), a correspondingly different expression for the solution (p{t) is obtained. EXAMPLE3.6. In (LN), \QtxeR^,to A(t).
= 0, and g(t)-
X(0):
0
The state transition matrix for A(f) has been determined in Example 3.4 as
Using (3.12), we obtain
W-e-^)
(l)h{t,to,xo)=^{t,0)x{0) using (3.11), we have
^(t,ri)g(ri)dri te
(i>p{t,tQ,XQ) = j and using (3.10), we finally obtain
(l)(t,to,Xo) = (l)h(t,to,Xo) + (l)p(t,to,Xo) =
'e-'{t-\) + y
D . H o w to D e t e r m i n e O(f,fo) To solve {LN) or {LN) in closed form, we require an expression for the state transition matrix O(f,fo)- We have seen in Example 3.5 that when A{t) is triangular, we can solve {LH) [and hence, O(f,fo)] by sequential integration of the individual differential equations. Eor the general case, however, closed-form determination of 0 ( f ,fo) is not possible. In the following, we identify another important class of matrices A{t) for which a closed-from expression of 0(f, t^) exists. THEOREM 3.8. If for every T,^ we have A(t)
I A{r\)dr\\ = JT
\
j \JT
A{n)dr\ A(t),
(3.14)
CO
Ul
^
'^(t,T) = eirMv)dv A ^ + y
then
_L
(3.15)
A(r])d7]
k=l
Proof, We recall that the general term of the Peano-Baker series [see Eq. (13.3) in Chapter 1] is given by rt
rsx
A{si)\
A{S2)"'
AySm) dSfn ClSfy,
'ds].
(3.16)
We wish to show that under the present assumptions, expression (3.16) is equal to the expression 1
A(s) ds
(3.17)
m! To verify (3.17), we will apply the identity A(r)
A(s) ds
dr =
|W+1
1 m+ 1
A(s) ds
(3.18)
repeatedly to (3.16). First, however, we verify the validity of (3.18). To accomplish this, we first observe that (3.18) is true for any fixed t = r. Differentiating the left-hand side of (3.18), we obtain d_ Jt
[Air,
m
I A(s)ds
= A(t) I A(s)ds
dr
(3.19)
Differentiating the right-hand side of (3.18), we have that -im+O
^\ dt
1 [ A{s)ds m+ 1 1 A{t) m+ 1
A{si)ds2'
A{sx)dsAA{i)
+
A{s\)ds\'
A{Sni^{)ds
I A{s^)ds^
A{Sm-^{)dSm^X +
A{Sm)dSm\A{t)\
A(s) ds
= A(t)
(3.20)
where in the last step of (3.20), the assumption (3.14) has been used repeatedly. Using (3.19) and (3.20), we obtain (3.18). To complete the proof of the theorem, we apply (3.18) repeatedly to (3.16) to obtain (3.17), the general term of the Peano-Baker series (13.13) in Chapter 1. Indeed, we have A(si)
A(s2y
A(Sm-1)
A(si)r
A(Sm) dSm dSm-1 dSm-2'
A(s2y
A(Sm-2)^ Sm-4
=
^j'A(si)pA(s2y'
r
"
dSi
A(s) ds
ds.'m-2'-'dSi
I
A(Sm-3):^
= ••• = -^W A(s)ds ml [J^ which was to be shown. This completes the proof of the theorem.
A(s) ds dStn-3'"dSi
CHAPTER 2 :
Response of Linear Systems
148 Linear Systems
For the scalar case, i.e., when A(t) = a(t), relation (3.14) is always true. Also, when A(t) = diag[aii(t)] [i.e., A(t) is a diagonal matrix], relation (3.14) is true. Furthermore, for A(t) = A, a constant matrix, (3.14) will always hold. The reader can readily verify that A(t) given in Example 3.1 also satisfies relation (3.14). We conclude by pointing out that for A G C[R, R''^''], (3.14) is true if and only if A(t)A(T) =
A(T)A(0
(3.21)
for all t and r. We ask the reader to verify the validity of this statement.
2.4 LINEAR SYSTEMS WITH CONSTANT COEFFICIENTS In this section we consider systems of linear, autonomous, homogeneous ordinary differential equations X =^ Ax
(L)
and systems of linear nonhomogeneous ordinary differential equations X = Ax + g(tl
(4.1)
where ;c G /?", A G i^^^", and g G C(R, R""). In the special case when A(t) = A, system (LH) reduces to system (L) and system (LN) reduces to system (4.1). Consequently, the results of Section 2.3 are applicable to (L) as well as to (LH) and to (4.1) as well as to (LN). However, because of the special nature of (L) and (4.1), more detailed information can be determined.
A. Some Properties of e^^ Let D = {(t, x) \ t E. R,x Ei R^}. In view of the results of Section 1.13, it follows that for every (to, XQ) G D , the unique solution of (L) is given by (l)(t, to, Xo) =
A*(f - ^o)*
^ +. = 12
Xo
^'
= *(f, to)xo = 0(r - to)xo = e^^'-'^ho,
(4.2)
where ^(t — to) = e^^^~^^^ denotes the state transition matrix for (L). [By writing ^(ty to) = ^(t - to), we are using a slight abuse of notation.] In arriving at (4.2) we invoked Theorem 10.9 of Chapter 1 in Section 1.13, to show that the sequence {)m}, where ^0 - Sm(t - to)xo,
(^m(r, fo, ^o) = .=1
(4.3)
^'
converges uniformly and absolutely as m ^ oo to the unique solution ct)(t, to, xo) of (L) given by (4.2) on compact subsets of R. In the process of arriving at this result, we also proved the following results.
THEOREM4.1. Let A be a constant nX n matrix (which may be real or complex) and let Sfn(t) denote the partial sum of matrices defined by m
Sm(t) k=l
.
149
(4.4)
k\
Then each element of the matrix Sm(t) converges absolutely and uniformly on any finite t interval (-a, a), a > 0, SLS m ^ oo. Furthermore, Sm(t) = ASm-i(t) = Sm-i(t)A, and thus, the limit of Sm(t) SLS t ^ oo is a C^ function on R. Moreover, this limit commutes with A. • In view of the above result, the following definition makes sense (see also Section 1.13). DEFINITION 4.1. Let A be a constant nX n matrix (which may be real or complex). We define e^^ to be the matrix °°
e^^ = 1 + ^
fk
LA'
k=l
k\
(4.5)
for any -co < ^ < oo, and we call e^^ a matrix exponential. We are now in a position to provide the following characterizations of e^^. THEOREM 4.2. Let / = RJQ ^ J, and let A be a given constant matrix for (L). Then (i) 0 ( 0 - e^^ is a fundamental matrix for all t < (ii) The state transition matrix for (L) is given by ^(t, to) <^(t - tol t G / . (iii) ^^^i^^^2 = ^A(fi+f2) for all tu t2 G / . (iv) Ae"^' = e^'A for all t G / . (v) {e^')-^ = ^-^^ for alU G 7. Proof. By (4.5) and Theorem 4.1 we have that (d/dt)[e'^^] = hm^^oo ASm(t) = hm^_^oo Sm(t)A = Ae^^ = e^^A. Therefore, 0 ( 0 = e^^ is a solution of the matrix equation 6 = AO. Next, observe that 0(0) = /. It follows from Theorem 3.3 that J^^[^^^ = ^tr(At) -^ Q fQ^ ^Y[ t G R. Therefore, by Theorem 3.4 0 ( 0 = e^^ is a fundamental matrix for (L). We have proved parts (i) and (iv). To prove (iii), we note that in view of Theorem 3.6(iii), we have for any ti,t2 ^ R that 0(^1, t2) = 0(^1, 0)0(0, t2). By Theorem 3.6(i) we see that 0(f, to) solves (L) with ^Oo, to) = 1' It was just proved that ^ ( 0 = ^^(^~^o) is also a solution. By uniqueness, it follows that 0(r, to) = e^^^'^^K For t = ti, to = —t2, we therefore obtain eMti+t2) = ^(^f^^ _^2) =: (^(t^)i^(-t2)-\ and for t = tiJo = 0, we have O(fi,0) = e^h = ^{ti). Also, for ^ = 0, ^ = -^2, we obtain 0(0, -^2) = e^^^ = 0(-^2)~^. Therefore, ^^(^1+^2) = ^^^^^^2 for all ti, t2 G R. Finally, to prove (ii), we note that by (iii) we have 0(^, to) = ^^(^~^o) = / + XI=i[(^ - to)'/k\]A'' = 0(f - to) is a fundamental matrix for (L) with O(^o, k) = L Therefore, it is a state transition matrix for (L). • We conclude this section by stating the solution of (4.1), (f){t, to, xo) = 0 ( r - to)xo +
0(^ - s)g(s) ds e^^'-'\g{s)ds
to
At
e-^'g{s)ds,
(4.6)
CHAPTER 2 :
Response of Linear Systems
150 Linear Systems
for all t Ei R.ln arriving at (4.6), we have used expression (13.8) of Chapter 1 and the fact that in the present case, 0(r, to) = ^^(^~^o)
B. How to Determine e At We begin by considering the specific case 0 0
A =
a 0
(4.7)
From (4.5) it follows immediately that e"^' = I ^tA
1 0
=
at 1
(4.8)
As another example, we consider Ai 0
0 A2
(4.9)
where Ai, A2 E R. Again, from (4.5) it follows that k
^At
^
0 k=l oht
0
k\
0 e^^'
(4.10)
Unfortunately, in general it is much more difficult to evaluate the matrix exponential than the preceding examples suggest. In the following, we consider several methods of evaluating e^K The infinite series method In this case we evaluate the partial sum Sm{t) (see Theorem 4.1) m
SM-1
^k
+ Y.T^' k\ k=l
for some fixed t, say, ti, and for m = 1, 2 , . . . until no significant changes occur in succeeding sums. This yields the matrix e^^^. This method works reasonably well if the smallest and largest eigenvalues of A are not widely separated. In the same spirit as above, we could use any of the vector differential solvers to solve X = Ax, using the natural basis for R^ as n linearly independent initial conditions [i.e., using as initial conditions the vectors ei = (1, 0 , . . . , 0)^, ^2 = ( 0 , 1 , 0 , . . . , 0)^, ...,en = ( 0 , . . . , 0,1)^] and observing that in view of (4.2), the resulting solutions are the columns of e^^ (with ^o = 0). EXAMPLE 4.1. There are cases when the definition of e^^ (in series form) directly produces a closed-form expression. This occurs for example when A^ = 0 for some k. In particular, if all the eigenvalues of A are at the origin, then A^ = 0 for some A: < n. In this case, only afinitenumber of terms in (4.5) will be nonzero and e^^ can be evaluated in closed form. This was precisely the case in (4.7). •
151
The similarity transformation method Let us consider the initial-value problem X = Ax,
CHAPTER 2 :
(4.11)
x(to) = xo,
let P be a real nXn nonsingular matrix, and consider the transformation x = Py, or equivalently, y = P~^x. Differentiating both sides with respect to t, we obtain y = p-^x = P~~^APy = Jy,y(to) = yo = P'^^o- The solution of the above equation is given by il^(t,to,yo) =
(4.12)
e'^'-'^^p-^xo.
Using (4.12) and x = Py, we obtain for the solution of (4.11) cl>(tJo.xo) =
(4.13)
Pe^^'-'^^p-^xo.
Now suppose that the similarity transformation P given above has been chosen in such a manner that / =
(4.14)
p-^AP
is in Jordan canonical form (see Subsection 2.20). We first consider the case when A has n linearly independent eigenvectors, say, v/, that correspond to the eigenvalues A/ (not necessarily distinct), / = 1 , . . . , n. (Necessary and sufficient conditions for this to be the case are given in Section 2.2. A sufficient condition for the eigenvectors Vi, i = 1 , . . . , ^, to be linearly independent is that the eigenvalues of A, A i , . . . , A„, be distinct.) Then P can be chosen so that P = [vi,..., v„] and the matrix / = P~^AP assumes the form 0 (4.15)
/ = 0
Afj
Using the power series representation 00
(4.16) k=i
we immediately obtain the expression ,Ai«
Jt
(4.17) 0
^A„r
Accordingly, the solution of the initial-value problem (4.11) is now given by rgAi(/-/o)
0
(t)(t, to, Xo) = P
P~'xo. 0
(4.18)
g,A„(r-?o)
In the general case when A has repeated eigenvalues, it is no longer possible to diagonalize A (see Subsection 2.2L). However, we can generate n linearly independent vectors v i , . . . , v„ and an n X n similarity transformation P = [vi,..., v„] that takes A into the Jordan canonical form / = P~^AP. Here / is in the block diagonal
Response of Linear Systems
152
form given by
Linear Systems
Jo Ji
/ =
(4.19)
0 where Jo is a diagonal matrix with diagonal elements A i , . . . , A^ (not necessarily distinct), and each 7,, / > 1, is an n, X m matrix of the form 'h+i 0
1 Ak+i
0 1
...
0 0 (4.20)
Ji =
0 0
0 0
... • 0 .
1 ^k+i
where X^+i need not be different from A,t+; if i ^ 7» and where k + ni-\ Now since for any square block diagonal matrix
h n^ = n.
c= 0 with Ci,i = I,..., I, square, we have that
cf
0
C' = 0 it follows from the power series representation of e-^' that \gJot
0 pJ\t
e'' =
(4.21) oJ st
t E. R. As shown earlier, we have Alt „Jot
0
=
(4.22) 0
For Ji, i = I,.. .,s,we
have /,• = Xk+ili + Ni,
(4.23)
where 7, denotes the n, X n, identity matrix and A^, is the «, X n, nilpotent matrix given by [O
1
...
Ol (4.24)
Ni :
0
••
1
0
153
Since \k+iU and Nt commute, we have that ^Jit
^
^Xk+it^Nit^
(4.25)
Repeated multiplication of Nt by itself results in Nf = 0 for all k^ the series defining e^^' terminates, resulting in
ni. Therefore,
fUi-l
\
t
. («,• -
0 1
o^k+it
JJi
0
1)!
/ = 1 , . . . , 5.
{ni - 2)!
0
(4.26)
1
It now follows that the solution of (4.11) is given by 0 0
0 0
4>it, to, xo) = P
pJ\(t-to)
P'^xo.
(4.27)
(fJs{t-to)
0 -1 0
EXAMPLE 4.2. In system (4.11), let A =
2 . The eigenvalues of A are A i = - 1 1
and A2 = 1, and corresponding eigenvectors for A are given by vi = (1,0)^ and V2 = (1, 1)^, respectively. Then P = [vi, V2] = -1 0
P-^AP = ^At
= Pe-^'P-^ =
1 0
1 1
2 1 1 0
1
-1 0
1 1 0
0
0
-1 1
1 1 ,P~' .0 1
0 1
Ai
0
0 A2
=
1 0
-1 and / = 1
as expected. We obtain
0
Suppose next that in (4.11) the matrix A is either in companion form or that it has been transformed into this form via some suitable similarity transformation P, so that A = Ac, where
0 0
1 0
0 1
0 0
0
0
0
1
(4.28)
Ar = -flO
-ai
-«2
Since in this case we have xt+i = Xi,i = 1 , . . . , n - 1, it should be clear that in the calculation of e"^^ we need to determine, via some method, only the first row of e^^. We demonstrate this by means of a specific example. EXAMPLE 4.3. In system (4.11), assume that A = Ac =
0 L-2
11 , which is in com-3J
panion form. To demonstrate the above observation, let us compute e^^ by some other method, say, diagonalization. The eigenvalues of A are Ai = - 1 and A2 = - 2 , and a set of corresponding eigenvectors is given by vi = (1, - 1 ) ^ and V2 = (1, - 2 ) ^ . We obtain
CHAPTER 2 :
Response of Linear Systems
154
,P-' =
P = [vi,V2] =
Linear Systems
1
2 -1
1 and J = p-^ArP = - 1 0 -1
0 At -2 , e
_
lip-^ ^ IM ^
i-2e-' + 2^-20 - 1 -2jL0 ^-^'JL-l - 1 . We note that the second row of the above matrix is the derivative of the first row, as expected. • The Cayley-Hamilton Theorem method If a(A) = det(\I - A) is the characteristic polynomial of an n X n matrix A, then in viev^ of the Cayley-Hamilton Theorem, we have that a(A) = 0, i.e., every nXn matrix satisfies its characteristic equation (refer to Subsection 2.2J). Using this result, along with the series definition of the matrix exponential 6^^ it is easily shown that n-l
e^' =
^ai{t)A\ i=0
[Refer to Subsection 2.2J for the details on how to determine the terms adt).] The Laplace transform method We assume that the reader is familiar with the basics of the (one-sided) Laplace transform. If / ( / ) = [fi(tl..., fn(t)f, where ft : [0,^)-^ R,i = I,..., n, and if each fi is Laplace transformable, then we define the Laplace transform of the vector/ componentwise, i.e., f{s) = [fi{s\ . ..,fn{s)Y, where fi{s) = X[Mt)] = We define the Laplace transform of a matrix C(t) = [cij(t)] similarly. Thus, if each ctj : [0, ^) —> R and if each c/y is Laplace transformable, then the Laplace transform of C(0 is defined as C(5') = iE[cij(t)] = [i£cij(t)] = [cij(s)]. Laplace transforms of some of the common time signals are enumerated in Table 4.1. Also, in Table 4.2 we summarize some of the more important properties of the Laplace transform. In Table 4.1, 8(t) denotes the Dirac delta distribution (see Subsection 1.16C) and pit) represents the unit step function. Now consider once more the initial-value problem (4.11), letting to = 0, i.e.. X = Ax,
TABLE 4.1
Laplace transforms
ma ^ 0) 8(t) Pit) t^lk\ ^-at fk^-at
e~"^ sin bt e~''^ cos bt
fis) = Wit)] 1 1/s 1/^^+1 l/(^ + a)
Wis + af^^ bilis + af + /?2] is + a)l\_is + of-\-b'^]
x(0) = XQ.
(4.29)
155
TABLE4.2 Laplace transform properties
CHAPTER 2 :
Time differentiation df{t)/dt d^f{t)/dt^ Frequency shift e--^f{t) Time shift f{t — a)p{t — a),a > 0 Scahng f{t/a),a>0 Convolution Jl>f(T)g(t-T)dT = f{t)*g{t) Initial value lim,^o+/W=/(0+) lim,^^/(f) Final value
*/(^)-/(o) ,*/(,) _ [ / - i / ( 0 ) + . .+fik- i)(0)]
/ > + «) e-"'f(s) af(as)
fim^)
linis^^ sf{sy lims^o sf{s)^''
^ If the limit exists. •'"'• If sf{s) has no singularities on the imaginary axis or in the right half s plane. Taking the Laplace transform of both sides of i = Ax, and taking into account the initial condition x(0) = XQ, we obtain sx{s) —xo= Ax{s), or {sI—A)x{s) = XQ, or x{s) = {sI-A)-^xo.
(4.30)
It can be shown by analytic continuation that {si — A)~^ exists for all s, except at the eigenvalues of A. Taking the inverse Laplace transform of (4.30), we obtain the solution 0 ( 0 = ^~^[{sI-A)-^]xo = O(f,0)jco = e^'xQ. (4.31) It follows from (4.29) and (4.31) that 6{s) = (sI-A)-^
and that
O(f,0)=O(f-0)=O(0=^"M(^/-A)-^]=/^
(4.32)
Finally, note that when to ^ 0, we can immediately compute 0 ( f ,fo) = ^ ( ^ — ^o) = ^ -1 0
EXAMPLE 4.4. In (4.29), let A
(sI-A)-
5+1 0
1 5+1
-2 s-l
2] . Then 1 2 (5+l)(5-l)
0
Using Table 4.1, we obtain i f ' ^
1 5+1
1 \S-l
0
s-l
1 5+1 1
{e'-e-')
[{si-A
0
Before concluding this subsection, we briefly consider initial-value problems determined by (4.1), i.e., x=Ax^g{t),
x{to)=xo.
(4.33)
We wish to apply the Laplace transform method discussed above in solving (4.33). To this end we assume ^o = 0 and we take the Laplace transform of both sides of (4.33) to obtain sx{s) —XQ =AX{S) -\-g{s) or {si — A)x{s) =xo-\-g{s), or x{s) = = ^
{sI-A)-\^{sI-A)-^g{s) ^{s)xo^^{s)g{s) ^h{s)^^p{s).
(4.34)
Response of Linear Systems
156 Linear Systems
Taking the inverse Laplace transform of both sides of (4.34) and using (4.6) with ^0 = 0, we obtain 0 ( 0 = 0/z(O + >p(0 = i^~^[(sl - A)~'^]xo + iE~^[(sI Ay^g(s)] = ^(t)xo + IQ ^(t ~ r])g(r])drj, where 0/^ denotes the homogeneous solution and (l)p is the particular solution, as expected. EXAMPLE4.5. Consider the initial-value problem given by Xi = — Xi + X2 X2 = - 2 ^ 2 +
U(t)
with xi(0) = -I, X2(0) = 0, and for t > 0, for t < 0.
u(t) = It is easily verified that in this case 1 / 1 s+1 s +I
1 s+2 1 s -h2
0
-It
0
e ^ (e ^ — e ^0' - 1 0. 0 e-' 1 5+ 1
/
1 5+1
2
1 -2t 2'
' 2
and
1 s+2
2 i J ' ^ 2\5 + 2
1 s+2
0 Ct>p{t) =
-e 0
2U
s+1
1/ 1 2 1^ + 2
_
2'
0 ( 0 = 0/,(O + 0p(O
2e- + i^-2^-
C. M o d e s a n d A s y m p t o t i c B e h a v i o r of Time-Invariant S y s t e m s In this subsection we study the qualitative behavior of the solutions of tonomous, homogeneous ordinary differential equations (L) by means of of such systems, to be introduced shortly. Although we will not address ity of systems in detail until Chapter 6, the results here will enable us to general stability characterizations for such systems.
linear, authe modes the stabilgive some
Modes: General case We begin by recalling that the unique solution of i: = Ax,
(L)
satisfying x(0) = JCQ, is given by (/>(r, 0, jco) = $ ( r , O)x(O) = 0 ( r , 0)xo = e'^'xQ.
(4.35)
We also recall that det(sl - A) = YlJ^i(s - KT\ where A i , . . . , A^- denote the a distinct eigenvalues of A, where A/ with / = 1 , . . . , cr, is assumed to be repeated nt times (i.e., nt is the algebraic multiplicity of A/), and Xf^im = n. To introduce the modes for (L), we must show that (T
flj-l
^At
= 1 ^=0
= J][Aioe^^' + Ante^^' + • • • + A,-(„,.where
1 Afk = -^-z
1 ^
-ix^t
j-r lim{[(s - XiT^sI - A)' i^im-i-k) }.
(4.36) (4.37)
In (4.37), [ • ]^^^ denotes the /th derivative with respect to s. Equation (4.36) shows that e"^^ can be expressed as the sum of terms of the form Aikt^e^'\ where A/^ E R^^^, We call Ai^t^e^'^ a mode of system (L). If an eigenvalue A/ is repeated nt times, there are nt modes, At^t^e^'^ k = 0, I,..., rii — \, in e^^ associated with A/. Accordingly, the solution (4.35) of (L) is determined by the n modes of (L) corresponding to the n eigenvalues of A and by the initial condition x(0). We note that by selecting x(0) appropriately, modes can be combined or eliminated [Aikx{Q) ^ 0], thus affecting the behavior of <^(t, 0, XQ). To verify (4.36) we recall that e^^ = i£~^[(sl - A)"^] and we make use of the partial fraction expansion method to determine the inverse Laplace transform. As in the scalar case, it can be shown that {si - A)-'
= X ^(klAaXs
- Xir^'^'\
(4.38)
i = l k^O
where the (klAfk) are the coefficients of the partial fractions (k\ is for scaling). It is known that these coefficients can be evaluated for each / by multiplying both sides of (4.38) by (s - \iY\ differentiating {nt - \ - k) times with respect to s, and then evaluating the resulting expression dXs — A/. This yields (4.37). Taking the inverse Laplace transform of (4.38) and using the fact that ie[^^^^^n = ^!(^-A/)"^^+^^ (refer to Table 4.1) results in (4.36). When all n eigenvalues kt of A are distinct, then cr = n, n/ = 1, / ^ 1 , . . . , n, and (4.36) reduces to the expression ^At
= Y.^ie M
(4.39)
i= l
where
lim[(^-A/)(^/-A)-^].
(4.40)
Expression (4.40) can also be derived directly, using a partial fraction expansion of (si - A)"^ given in (4.38) (verify this). EXAMPLE 4.6. For (L) we let A
0
1] , for which the eigenvalue Ai
-4 - 4 J
repeated twice, i.e., ni = 2. Applying (4.36) and (4.37), we obtain 1 0 e-^^ + Aioe^^^ ^ Anteht te 0 1
-2 is
157 CHAPTER!:
Response of Linear Systems
158 Linear Systems
0 for which the eigenvalues are given -1 by (the complex conjugate pair) Ai = _i2 + j ( V3/2), A2 = - i - 7( V3/2). Applying (4.39) and (4.40), we obtain EXAMPLE 4.7. For (L) we let A =
1 A^ =
1
Ai + 1
Ai - A2
^J~3 A2 + 1 -1
A2 =
.73
.
1
.73
-"2 + ^-2
7^
_1
1
1
-1 2
^~2
-J 73
-1
1
7^
2
-^ 2 J
[i.e., Ai = A2, where (• )* denotes the complex conjugate of (•)], and e^^ = Aie^^' + A2e^^' = Aie^^' + A\e^*i' = 2(Re Ai)(Re e^^') -
2(ImAi)(Ime^'')
1 = 2e-^"^^'
0
2 0
75
1
-:
273 1
-
COS -—t
2J
1
•
~7i sinM.j 1
[ 7^ 273.
\
/
The last expression involves only real numbers, as expected, since A and e^^ are real matrices. • ri 01 EXAMPLE 4.8. For (L) we let A = for which the eigenvalue Ai = 1 is re[0 i j peated twice, i.e., n\ = 2. Applying (4.36) and (4.37), we obtain 1 0
Aioe^^^ + Aute^^^
0 0 e' + 1 0
0 te' = Ie\ 0
This example shows that not all modes of the system are necessarily present in e^K What is present depends in fact on the number and dimensions of the individual blocks of the Jordan canonical form of A corresponding to identical eigenvalues. To illustrate this further, we let for (L), A =
0
1
, where the two repeated eigenvalues Ai = 1
belong to the same Jordan block. Then e^
1 0
0 0 e' + 1 0
1 te'. 0
Stability of an equilibrium In Chapter 6 we will study the qualitative properties of linear dynamical systems, including systems described by (L). This will be accomplished by studying the stability properties of such systems, or more specifically, the stability properties of an equilibrium of such systems. If 0(/, 0, Xe) denotes the solution of system (L) with x(0) == Xe, then Xe is said to be an equilibrium of (L) if (j){t, 0, Xe) ^ Xe for all f > 0. Clearly, jc^ = 0 is an equilibrium of (L). In discussing the qualitative properties, it is often customary to speak, somewhat loosely, of the stability properties of system (L), rather than the stability properties of the equilibrium x^ = 0 of system (L).
We will show in Chapter 6 that the following qualitative characterizations of system (L) are actually equivalent to more fundamental qualitative characterizations of the equilibrium Xe = 0 of system (L):
159 CHAPTER 2 :
Response of Linear Systems
1. The system (L) is said to be stable if all solutions of (L) are bounded for all t > 0 [i.e., for any solution (f){t, 0, XQ) = (4>i{t, 0, XQ), ..., 4>n{h 0. -^o))^ of i^)^ there exist constants Mi, i = I,.. .,n (which in general will depend on the solution on hand) such that \(f)i{t, 0, xo)| < M/ for all t > 0]. 2. The system (L) is said to be asymptotically stable if it is stable and if all solutions of (L) tend to the origin as t tends to infinity [i.e., for any solution 4>{t, 0, XQ) = {4>\{t, 0, xo),..., (l)n{t, 0, xo)Y of (L), we have lim^_^oo 4>i{t, 0, xo) = 0,i = 1,...,^]. 3. The system (L) is said to be unstable if it is not stable. By inspecting the modes of (L) given by (4.36), (4.37) and (4.39), (4.40), the following stability criteria for system (L) are now evident: 1. The system (L) is asymptotically stable if and only if all eigenvalues of A have negative real parts (i.e.. Re Xj <0, j = 1 , . . . , a). 2. The system (L) is stable if and only if Re A^ < 0, j = 1 , . . . , a, and for all eigenvalues with Re Xj = 0 having multiplicity HJ > I, it is true that lim [(s
k=
I,
1. (4.41)
3. System (L) is unstable if and only if (2) is not true. We note in particular that if Re Xj = 0 and nj > 1, then there will be modes ^jkl^y k = 0,.. .,nj - 1, that will yield terms in (4.36) whose norm will tend to infinity as ^ ^ oo, unless their coefficients are zero. This shows why the necessary and sufficient conditions for stability of (L) include condition (4.41). EXAMPLE 4.9. The systems in Examples 4.6 and 4.7 are asymptotically stable. A system (L) with A
ro [0
n is stable, since the eigenvalues of A above are Ai
-1
=
- 1 01 is unstable since the eigenvalues of A are 0 1 -1. The system of Example 4.8 is also unstable.
0, A2 = - 1 . A system (L) with A 1,A2
Modes: Distinct eigenvalue case When the eigenvalues A/ of A are distinct, there is an alternative way to (4.40) of computing the matrix coefficients A/, expressed in terms of the corresponding right and left eigenvectors of A. This method offers great insight in questions concerning the presence or absence of modes in the response of a system. Specifically, if A has n distinct eigenvalues A^, then ^At
where
-X^i Ai
ViVi,
M
(4.42) (4.43)
where vi G R^ and {viY E R^ are right and left eigenvectors of A corresponding to the eigenvalue A/, respectively.
160 Linear Systems
To prove the above assertions, we recall that ( A / / - A)v/ = 0 and v ^ A / / - A) = O.lf Q= [vi,..., Vn], then the v/ are the rows of
The matrix Q is of course nonsingular, since the eigenvalues A/, / = 1 , . . . , n, are by assumption distinct and since the corresponding eigenvectors are linearly independent. Notice that Qdiag[k\,.. .,Xn\ = ^Q and that diag[k\,..., A„]P = PA. Also, notice that v/Vj = 5/y, where 8ij =
1 0
when / = j , when / T^ j .
We now have ( ^ / - A ) - ^ = [si-Qdiag [\u ,.., XnlQ'^T = Q[sl-diag[Xu ..., K]r'Q-' - Qdiag[(s-X,r\...,(s-Xnr']Q-' = Sf=iV,vK^-AO-i.Ifwe now take the inverse Laplace transform of the above expression, we obtain (4.42). If we choose the initial value x(0) for (L) to be cohnear with an eigenvector vj of A [i.e., x(0) = avj for some real a ¥= 0], then e^j^ is the only mode that will appear in the solution cf) of (L). This can easily be seen from our preceding discussion. In particular if x(0) = aVj, then (4.42) and (4.43) yield ^"' - avie^J' (/>(^, 0, x(0)) = e'^'xiO) = vivi40)^^^' + • • • + VnVnX(0)e
(4.44)
since v/Vy = 1 when / = y, and v/Vy = 0 otherwise. EXAMPLE 4.10. In (L) we let A = Ai = - 1 , A2 = 1 and Q = [vi,V2]
r-1 0 1 0
1 . The eigenvalues of A are given by 1
1 -i . Then e"^'
1 2 M~
0
1 -i ^Mfin particular we choose x(0) = avi = 0 0 (a, 0)^, then (pit, 0, x(0)) = e^^x(0) = a(l, O^e'^ which contains only the mode corresponding to the eigenvalue Ai = - 1 . Thus, for this particular choice of initial vector, the unstable behavior of the system is suppressed. • v\V\e^^^ + V2V2e^'^^ =
Remark We conclude our discussion of modes and asymptotic behavior by briefly considering systems of linear, nonhomogeneous, ordinary differential equations (4.1) for the special case where g(t) = Bu{t), X = Ax^
Bu(t\
(4.45)
where B G R^^^^ ^ - R ^ Rm ^^^ where it is assumed that the Laplace transform of u exists. Taking the Laplace transform of both sides of (4.45) and rearranging yields x{s) = (si - A)-'x(0)
+ (si -
A)~'Bu(s).
(4.46)
By taking the inverse Laplace transform of (4.46), we see that the solution cf) is the sum of modes that correspond to the singularities or poles of (si - A)~ ^ x(0) and (si A)~^Bu(s). If in particular (L) is asymptotically stable (i.e., for x = Ax, Re Xi <
0, / = 1,.. .,n) and if u in (4.45) is bounded (i.e., there is an M such that 1^^(01 < M for all r > 0, / = 1 , . . . , m), then it is easily seen that the solutions of (4.45) are bounded as well. Thus, the fact that the system (L) is asymptotically stable has repercussions on the asymptotic behavior of the solution of (4.45). Issues of this type will be addressed in greater detail in Chapter 6.
*2.5 LINEAR PERIODIC SYSTEMS We now consider linear homogeneous systems of first-order ordinary differential equations X = A(t)x,
-co
(P)
where A G C(R, /?"^") and A(t) = A(t + T),
-00 < ^ < 00,
(5.1)
for some T > 0. We call (P) a. periodic system and T a. period for system (P). The principal result of this section involves the notion of the logarithm of a matrix, which we introduce in the following result. THEOREM 5.1. For every nonsingular matrix B there exists a matrix A, called a logarithm ofB, with the property that (5.2)
B. The matrix A is not unique.
Proof. Let B be similar to B. Then there exists a nonsingular matrix P such that P-^BP = B. Now if e^ = B, then we have B = PBP'^ = Pe^P'^ = e^^^~\ It follows that PAP'^ is also a logarithm of 5. Therefore, it suffices to prove the theorem when the matrix B is in suitable canonical form. Let Ai,..., Ayt denote the distinct eigenvalues of B with respective multiplicities ni,.. .,nk. Without loss of generality, we may assume that B is in the block diagonal form r^i 0 B= 0
Hj
where Bj = Ay[/„. + (1/\J)NJIN''/ = OJ = h..., k. We note that A^ T^ 0, y = l,...,k, since B is nonsingular. Using the power series expansion log(l + x) = Z^p=i[(-iy^^/p]xP,\x\ < 1, we formally write A^ = logBj = /„.logAy + log[/,. + (l/\j)Nj] = InjlogXj + X;=d(-^y^'^PWAjy> since 7V7 = 0 we actually have
Aj = /„^logA, + X L ^ / ^ T ,
j = i,„„k,
(5.3)
where we note that logA^ is defined, since Xj 7^ 0. Now recall that e'°8(i+j:) = I + x. Performing the same operations with matrices, we obtain the same terms and there
161 CHAPTER 2:
Response of Linear Systems
162 Linear Systems
is no problem with convergence, since the series (5.3) for Aj = ^ogBj terminates. Accordingly, we obtain e^J = exp(Inj log\j)Qxp{Z.''J~^[(-l)P^^/p](Nj/Xj)P} = \j[Inj + (Nj/\j)] = Bj, j = 1,..., ^. If now we let A= 0
where Aj is defined in (5.3), we obtain .A
e^i
0
0
e^k
^
= B, 0
Bk
which is the desired result. We conclude by noting that the matrix A is not unique, since for example, e^^^'^^^^ = ^A^iTTj ^ ^A f^j. ^u integers k (where j = v - 1 ) . • We are now in a position to state and prove one of the principal results of this section. THEOREM 5.2. Assume that (5.1) is true and that A E C{R, /?">^"). If <|)(0 is a fundamental matrix for (P), then so is ^{t +T),t EL R . Furthermore, for every ^ there exists a nonsingular matrix P that is also periodic with period T and a constant nXn matrix R, such that ^ ( 0 = P{t)e'^. Proof. Let ^ ( 0 = ^(? + T\t E R. Since i>(0 = A(t)^(t\ t E R, we have ^ ( 0 = 4)(^ + 7) = A(t + T)^(t + r ) = A{t)(t + T),tE R. Therefore, ^ is also a solution of ^ = A(0^, A(t) = A{t + T\t E R. Furthermore, since ^{t + T) is nonsingular for all t E R/ii follows that ^ is a fundamental matrix for (P). Therefore, by Theorem 2.5, there exists a nonsingular matrix C such that ^{t + T) = ^{t)C, and by Theorem 5.1, there exists a constant matrix R such that e^^ = C. Therefore, ^(t + T) = ^(t)e TR
(5.4)
P(t) = ^(0^ -tR
(5.5)
Defining P by
and using (5.4) and (5.5), we now obtain P{t + T) = ^(r + T)e-^^-^^^^ = ^{t)e^^ X ^-(t+T)R ^ (^(t)e-tR = p{t). Therefore, P{t) is nonsingular for dll t E R and it is periodic. • The above result allows us to conclude that the determination of a fundamental matrix O for system (P) over any time interval of length T leads at once to the determination o/O over (-oo, oo). To show this, assume that ^(t) is known only over the interval [^o. to> -^T], Since ^(t + T) = ^(t)C, we obtain by setting t = to,C = 0(^)~iO(ro + T) and R = T~^ log C. It follows that P(t) - $(0^~^^ is now also known over [to, to -\- T]. However, P(t) is periodic over (-00,00). Therefore, 4>(0 is givenover (-00, 00) by $ ( 0 = P(t)e^^. Next, let 0 be any other fundamental matrix for (P) with A(t) = A(t + T). Then O = $ 5 for some constant nonsingular matrix S. Since ^(t + T) = ^(t)e^^,
163
we have that ^(t + T)S = Mt)Se'^^, or (5.6) This shows that every fundamental matrix O of(P) determines a matrix Se^^S~^ that is similar to the matrix e^^. Conversely, let S be any constant nonsingular matrix. Then there exists a fundamental matrix of (P) such that Eq. (5.6) holds. Therefore, even though $ does not determine R uniquely, the set of all fundamental matrices of (P), and hence of A, determines uniquely all parameters associated with e^^ that are invariant under a similarity transformation. In particular the set of all fundamental matrices of A determines a unique set of eigenvalues of the matrix e^^, denoted by A i , . . . , A„, that are called the Floquet multipliers associated with A. Note that none of these vanish, since nf=i A/ = dete^^ # 0. The eigenvalues of P are called the c/z^rac^^mr/c ^x/76>/i^n/5'. Next, we let 2 be a constant nonsingular matrix that transforms R into its Jordan canonical form, i.e., / ^ Q~^RQ, where Jo 0
0 /i
0
0
/ =
Now let O = Q and let P = PQ. In view of Theorem 5.2, we have that ^ ( 0 - P{t)e'\
Pit) = Pit + T).
(5.7)
Let the eigenvalues of R be denoted by p i , . . . , p^. Then
JJ
where
O^JO
=
0
e'^'
0
0
0 0 ptJs
0 0
e'P' 0 0 e'P^ 0
0
e'Pi
and
jJi
_
Jpq^i
'
^
2
0
1
t
0
0
0
(n - 1)! fn-2
in - 2)!
i =
l,...,s,
q + ^rt
-= n.
i=l
1
Now A; = e^P'. Therefore, even though the p, are not uniquely determined, their real parts are. It follows from (5.7) that the columns 4>\>--->4>n of "J* are linearly independent solutions of {P). Let pi,..., p„ denote the periodic column vectors
CHAPTER 2 :
Response of Linear Systems
164 Linear Systems
of P. Then 01 (0 = e^'^piit) hit)
= e'P^p2(t)
4>q{t) = e'P^Pqit) 4>q+\{t) = e'P^-'pg+i(t) (5.8)
4>q+2(t) = e'P^-'(tPg+i(t) + Pq+2(t))
^q-,n(t) = e''^-^ (^ - iyPq+l(t) + ••• + tpg+r,-l(t) + Pq+r,(t)
4 - r , + l ( 0 = e'P^-^ Pn-rs + l(t\
fVs-l
4>n(t)
otPq+s
(rs - 1)!
Pn-rs + l(0 + • • • + tpn-\{t)
+ pnit)
From (5.8) it is easy to see that when Re pt — at < 0, or equivalently, when |A/| < 1, there exists a A:, > 0 such that \h(t)\ < kie^'^^''^^^ -> 0 as f ^ oo. This shows that if the eigenvalues pu i = I,..., n, of R have negative real parts, then any norm of any solution of (P) tends to zero as ^ -^ -\-^ at an exponential rate. From (5.5), we can easily verify by direct computation that AP - P = PR. Accordingly, for the transformation X = Pit)y,
(5.9)
we obtain x = A(t)x = A(t)P(t)y = P(t)y + P(t)y = (d/dt)[P(t)y] or j = P~\t) X [A(t)P(t) - P(t)]y = p-\t)[P(t)R]y = Ry. In other words, the transformation (5.9) reduces the linear, homogeneous, periodic system (P) to the system y = Ry, a linear homogeneous system with constant coefficients. We conclude this section with a specific example. EXAMPLES.1. Consider the scalar system X = -(sin/ + 2)x
(5.10)
Then A(t) = -(sin/ + 2), and A(t) is periodic with period T = ITT. A fundamental matrix for (5.10) is given by 0(0 = exp(cosr - 1 - 2/) as can be verified by substituting into the relation i>(0 = A(t)^(t). Letting t = 0 and T = ITT in (5.4), we obtain 0(27r) - ^"^^ = 0(0)^^''^ = e^""^ or R = - 2 . The equivalence matrix P(t) is now given by (5.5) as P{t) = exp (cos t - l - 2t)e^^ = e^^^^~^, which is clearly periodic with period T = ITT. The given system (5.10) is transformed by P{t) into the system y = (^i-^«^0[(-l)(sin/ + 2)^'^°^^-i + smte''''''-^]y = -(e^-^^'')(2e'''''-^)y = -2y = Ry. We will address some of the qualitative properties of periodic systems in further detail in Chapter 6.
165
2.6 STATE EQUATION AND INPUT-OUTPUT DESCRIPTION OF CONTINUOUS-TIME SYSTEMS
CHAPTER 2 :
Response of Linear Systems
This section consists of three subsections. Using the material of the preceeding sections of this chapter, we first study the response of Hnear continuous-time systems. Next, we examine transfer functions of Unear time-invariant systems, given the state equations of such systems. Finally, we explore the equivalence of internal representations of systems.
A. Response of Linear Continuous-Time Systems Returning now to Sections 1.1 and 1.14, we consider once more systems described by linear time-varying equations of the form X = A(t)x -h B(t)u
(6.1a)
y = C(t)x -h D(t)u,
(6.1b)
where A G C(R, R'"'"''), B G C(R, /?^><^), C G C(R, RP''''\ D G C(R, RP"""^), and u : R-^ R^ is assumed to be continuous or piecewise continuous. We recall that in (6.1a) and (6.1b), x denotes the state vector, u denotes the system input, and y denotes the system output. From Section 1.14 we recall that for given initial conditions ^0 ^ R> ^(to) = XQ E: R^ and for a given input u, the unique solution of (6.1a) is given by (l)(t, to, xo) = 0(r, to)xo +
^(t,
s)Bis)u(s)ds
(6.2)
for t E: R, where O denotes the state transition matrix of A(0. Furthermore, by substituting (6.2) into (6.1b), we obtain [as in (14.6) of Chapter 1], for all t G R, the total system response given by y(t) = C(t)^(t, to)xo + C(t)
0(^, s)B(s)u(s)ds + D(t)u(t).
(6.3)
Jto
Recall that the total response (6.3) may be viewed as consisting of the sum of two components, the zero-input response given by the term il/(t, to, Xo, 0) = C(t)<^(t, to)xo
(6.4)
and the zero-state response given by the term p(t, to, 0, u) = C{t)
^{t, s)B(s)u(s)ds -h D(t)u(t).
(6.5)
Jto
The cause of the former is the initial condition xo [and can be obtained from (6.3) by letting u(t) = 0], while for the latter the cause is the input u [and can be obtained by setting XQ = 0 in (6.3)]. The zero-state response can be used to introduce the impulse response of the system (6.1a), (6.1b). Returning to Subsection 1.16C, we recall that by using the
166
Dirac delta distribution S, we can rewrite (6.3) with XQ = 0 as
Linear Systems y(t) = f [C(tmt,
T)B(T)
+ D(t)8(t - T)]u(T)dT
Jto
= [ H(t,T)u(T)dT,
(6.6)
Jto
where H(t, r) denotes the impulse response matrix of system (6.1a), (6.1b) given by [ C(t)<^(h
H{t, T)
T)B{T)
+ D{t)8{t - r),
t^
[0,
T,
t< T.
(6.7)
When in (6.1a), (6.1b), A{t) = A, B(t) = B, C(t) = C, and D(t) = D, we obtain the time-invariant system (6.8a) (6.8b)
X = Ax + Bu y = Cx -^ Du. We recall that in this case the solution of (6.8a) is given by
(6.9) the total response of system (6.8a), (6.8b) is given by y{t) = Ce'^^'-'^ho + C [ e^^'-'^Bu{s)ds + Du{t\
(6.10)
and the zero-state response of (6.8a), (6.8b) is given by ^(0 = /^^[C^^^^~'^^B+ D8(t - T)]u(T)dT - \l H{t,T)u{T)dT = 1^ H(t - T)u(T)dT, where the impulse response matrix H of system (6.8a), (6.8b) is given by H(t -T)
=
Ce^^'-^^B + D8(t - r), 0,
t^ r, t
(6.11)
or, as is more commonly written, H(t) =
Ce^'B + D8{t\ 0,
t> 0, ^<0.
(6.12)
At this point it may be worthwhile to consider some specific cases. EXAMPLE 6.1. In (6.1a), (6.1b), let A{t) =
-1
0
^2n €'
-1
B{t) =
0
C(0 = [e\ 1],
D = 0
and consider the case when t^ = 0, x(0) = (0, 1)^, u is the unit step function, and r > 0. Referring to Example 3.6, we obtain (p(t, to, XQ) = (f)h{t, t^, XQ) + (j)p{t, to, xo) = 2 {e'
- €-')
te-'
0
with ^ = 0 and for ^ > 0. The total system response y(t) =
C(t)x(t) is given by the sum of the zero-input response and the zero-state response, y(t, to, Xo, u) = il/(t, to, Xo, 0) + p(t, to, 0, u) = [\{e'^^ - 1) + e~^] + t,t> ^. Note that the zero-input response ip is due to the homogeneous part of the solution 4> (given by (f)h) while the zero-state response p is due to the particular solution of (/> (given by (/)p).
EXAMPLE 6.2. In (6.8a), (6.8b), let A
C = [0,1], D = 0 and
[o o} ^ - [y
CHAPTER 2 :
consider the case when to = 0, x(0) = (1, - 1 ) ^ , u is the unit step, and t > 0. We can easily compute the solution of (6.8a) as (/)(^, to, Xo) = (t)h(t, to, Xo) + (l>p(t, to, Xo)
=
n -^1
^t^
+
-1
with ^0 = 0 and for ? > 0. The total system response y(t) = C(t)x(t) is given by the sum of the zero-input response and the zero-state response, y(t, to, xo, u) = il/(t, to, xo, 0) + p(t, to,0,u)=-l-\-t,t>0. • We note that when x(0) = 0, Example 6.1 (a time-varying system) and Example 6.2 (a time-invariant system) have identical output responses given by y(t) = t, r > 0, when u(t) is the unit step. [Is this true for any input u(t)l] EXAMPLE6.3. Consider the time-varying system given above in Example 6.1. In this case we have 0(?, T)B(T) = [e~\ 0]^, and the impulse response has the rather unusual form Hit, T) =
C{t)^{t,
T)B(T)
= 1,
0,
t^
T,
t
In other words, the response of this system to an impulse input, for zero initial conditions, is the unit step, and this is independent of the time r at which the impulse is applied! Note that in the present case the response to a step is a ramp t, as can easily be verified from (6.6) (see also Example 6.1). Therefore, this system behaves to the outside world, for zero initial conditions, as a time-invariant system. This is interesting; however, it is not a typical situation when dealing with time-varying systems. • EXAMPLE 6.4. Consider the time-invariant system given above in Example 6.2. It is easily verified that in the present case (^(t) = e^' =
n ^1 0
167
1
Then H(t, r) = Ce'^^'-^'^B = I for t > r and H{t, r) = OfoYt< r. Thus, the response of this system to an impulse input for zero initial conditions is the unit step. Comparing this with the impulse response of the system given above, in Example 6.3, we note that they are identical. In other words, the behavior of these two systems to the outside world, one a time-varying system and the other a time-invariant system, is characterized by the same response to an impulse input, when the initial conditions are zeros. Indeed, in this case, both systems behave like a time-invariant system with H(t, r) = H(t - r, 0) = H(t, 0) = 1. Note, however, that when the initial conditions are not zero, the responses of these two systems are quite different. • The preceding two examples demonstrate, as one might expect, that external descriptions of finite-dimensional linear systems are not as complete as internal descriptions of such systems. Indeed, the utility of impulse responses is found in the fact that they represent the input-output relations of a system quite well, assuming that the system is at rest. To describe other dynamic behavior, one needs in general additional information [e.g., the initial state vector (or perhaps the past history of the system input since the last time instant when the system was at rest) as well as the internal structure of the system]. Internal descriptions, such as state-space representations, constitute more complete descriptions than external descriptions. However, the latter are simpler to apply
Response of Linear Systems
168 Linear Systems
than the former. Both types of representations are useful. It is quite straightforward to obtain external descriptions of systems from internal descriptions, as was demonstrated in this section. The reverse process, however, is not quite as straightforward. The process of determining an internal system description from an external description is called realization and will be addressed in Chapter 5. The principal issue in system realization is to obtain minimal order internal descriptions that model a given system, avoiding the generation of unnecessary dynamics.
B. Transfer Functions Next, if as in (16.51) in Chapter 1, we take the Laplace transform of both sides of (6.12), we obtain the input-output relation y{s)=H{s)u{s).
(6.13)
We recall from Section 1.16 that H{s) is called the transfer function matrix of system (6.8a), (6.8b). We can evaluate this matrix in a straightforward manner by first taking the Laplace transform of both sides of (6.8a) and (6.8b) to obtain sx{s) - x ( 0 ) =Ax{s) +Bu{s)
(6.14)
y{s) = Cx{s) +Du{s).
(6.15)
Using (6.14) to solve foxx{s), we obtain x{s) = {sI-A)-^x(Qi) +
{sI-A)-^Bu{s).
(6.16)
Substituting (6.16) into (6.15) yields y{s) = C{sl-A)-^x{0) and
y{t) = ^~^y(s)
+Du{s)
(6.17)
e^^'-'^Bu(s)ds^Du(t),
(6.18)
^C{sI-A)-^Bu{s)
= C/'jc(0) + C / Jo
as expected. If in (6.17) we let x(0) = 0, we obtain the Laplace transform of the zero-state response given by y{s) =
[C{sI-A)-^B^D]u{s)
= H{s)u{s),
(6.19)
where H{s) denotes the transfer function of system (6.8a), (6.8b), given by H{s)=C{sI-A)-^B^D.
(6.20)
Recalling that ^[e^^] = O(^) = {si—A) ^ [refer to (4.32)], we could of course have obtained (6.20) directly by taking the Laplace transform of H{t) given in (6.12). EXAMPLE 6.5. In Example 6.2, let ^o = 0 and x(0) = 0. Then H{s) = C{sI-A)-^B + D = [0,1] ' "
^[0,1]
and H{t) = ^
rl s
l1 1 s
^H{s) = 1 for ^ > 0, as expected (see Example 6.2).
Next, as in Example 6.2, let x(0) = (1,-1)^ and let u be the unit step. Then y(s) = C(sl - A)-ix(O) + H(s)u(s) = [0, 1/^](1, -1)^ + (l/s)(l/s) = -1/s + 1/s^ and y(t) = ^~^[y(s)] = - 1 + ^ for / > 0, as expected (see Example 6.2). • We note that the eigenvalues of the matrix A in Example 6.5 are the roots of the equation det (si - A) = s'^ = 0, and are given by ^-i = 0, ^2 = 0, while the transfer function H(s) in this example has only one pole (the zero of its denominator polynomial), located at the origin. It will be shown in Chapter 5 (on realization) that the poles of the transfer function H(s) (of a SISO system) are in general a subset of the eigenvalues of A. In Chapter 3 we will introduce and study two important system theoretic concepts, called controllability and observability. We will show in Chapter 5 that the eigenvalues of A are precisely the poles of the transfer function H(s) = C(sl - Ay^B + D if and only if the system (6.8a), (6.8b) is observable and controllable. This is demonstrated in the next example. 0
0' C = [-3,3],/) = 0. ,5 = -2^ [-1 The eigenvalues ofA are the roots of the equation J^r (5/-A) = s^ + 2s+l = (^+1)^ = 0 given by ^i = - 1 , 5*2 = - 1 , and the transfer function of this SISO system is given by
EXAMPLE 6.6. In (6.8a), (6.8b), let A =
H(s) = C(sl-A)-^B
+D =
r
w
_ . - 1 .-11 [-3,3]^s 1 ^+ 2
3(s - 1) 5+ 2 1 (S+ 1 ) 2 ' -1 s with poles (the zeros of the denominator polynomial) also given by si = 3[-l,l]
1
(S + 1)2
If in Example 6.6 we replace B = [0, 1]^ and D = 0 by B =
-1,^2 = - 1 . and
D = [0, 0], then we have a multi-input system whose transfer function is given by H(s) =
'3(^-1) 3 (S+ 1)2^(^+1)
The concepts of poles and zeros for MIMO systems (also called multivariable systems) will be introduced in Chapter 4. The determination of the poles of such systems is not as straightforward as in the case of SISO systems. It turns out that in the present case the poles of H(s) axe si = -1,S2 = - 1 , the same as the eigenvalues of A. Before proceeding to our next topic, the equivalence of internal representations, an observation concerning the transfer function H(s) of system (6.8a), (6.8b), given by (6.20), H(s) = C(sl — A)~^B + D is in order. Since the numerator matrix polynomial of (si — A)"^ is of degree (n - 1) (refer to Subsection 2.1G), while its denominator polynomial, the characteristic polynomial a(s) of A, is of degree n, it is clear that \im H(s) = D, a real-valued mXn matrix, and in particular, when the ''direct link matrix" D in the output equation (6.8b) is zero, then lim^(^) = 0,
169 CHAPTER 2 :
Response of Linear Systems
170 Linear Systems
the m X n matrix with zeros as its entries. In the former case (when D 7^ OorD = 0), ^(^) i^ ^^^^ ^^ ^^ di proper transfer function, while in the latter case (when D = 0), H(s) is said to be a strictly proper transfer function. When discussing the realization of transfer functions by state-space descriptions (in Chapter 5), we will study the properties of transfer functions in greater detail. In this connection, we will also encounter systems that can be described by models corresponding to transfer functions H(s) that are not proper The differential equation representation of a differentiator (or an inductor) given by y(t) = (d/dt)u{t) is one such example. Indeed, in this case the system cannot be represented by Eqs. (6.8a), (6.8b) and the transfer function, given by H(s) = s is not proper. Such systems will be discussed in Chapter 7.
C. Equivalence of Internal Representations In Subsection 2.4B it was shown that when a linear, autonomous, homogeneous system of first-order ordinary differential equations i = Ax is subjected to an appropriately chosen similarity transformation, the resulting set of equations may be considerably easier to use and may exhibit latent properties of the system of equations. It is therefore natural that we consider a similar course of action in the case of the linear systems (6.1a), (6.1b) and (6.8a), (6.8b). We begin by considering (6.8a), (6.8b) first, letting X = Px,
(6.21)
where P is a real, nonsingular matrix (i.e., P is a similarity transformation). Consistent with what has been said thus far, we see that such transformations bring about a change of basis for the state space of system (6.8a), (6.8b). Application of (6.21) to this system will result, as will be seen, in a system description of the same form as (6.8a), (6.8b), but involving different state variables. We will say that the system (6.8a), (6.8b), and the system obtained by subjecting (6.8a), (6.8b) to the transformation (6.21), constitute equivalent internal representations of an underlying system. We will show that equivalent internal representations (of the same system) possess identical external descriptions, as one would expect, by showing that they have identical impulse responses and transfer function matrices. In connection with this discussion, two important notions called zero-input equivalence and zero-state equivalence of a system will arise in a natural manner. If we differentiate both sides of (6.21), and if we apply x = P~^x to (6.8a), (6.8b), we obtain the equivalent internal representation of (6.8a), (6.8b) given by k =^ Ax + Bu where
A = PAP-\
y = Cx + Du, B = PB, C = CP-\
(6.22a) D = D
(6.22b) (6.23)
and where x is given by (6.21). It is now easily verified that the system (6.8a), (6.8b) and the system (6.22a), (6.22b) have the same external representation. Recall that for (6.8a), (6.8b) and for (6.22a), (6.22b), we have for the impulse response H(t, T) ^ Hit - T, 0) = { ^'"""'^
+ ^^(' - ^)'
; J ;>
(6.24)
C^ia-T)^ + D8(t - T), 0, Recalling from Subsection 2.4B [see Eq. (4.13)] that and
H{t, r) = H{t - r, 0)
^A(r-T) ^
t < T.
(6.25)
T)
p^A(t-T)p-l
- H(t,
(6.26)
(6.27)
T),
and this in turn shows that (6.28)
His) = H(s\
This last relationship can also be verified by observing that^(^) == C(sl - A) ^ X B -{- D = CP-\sI - PAP-^)-^PB + D = CP-^P(sI - A)-^p-^PB + D = C(sl - A)-^B + D - H(s). Next, recall that in view of (6.10) we have for (6.8a), (6.8b) that y(t) = Ce^^'-'^^xo +\
Hit-
T,0)u(T)dT
J to
= il/(t, to, XQ, 0) + p(t, to, 0, u)
(6.29)
and for (6.22a), (6.22b) that y(t) = Ce^^'-'^ho +
H(t-T,0)u(T)dT (6.30)
= {{/(t, to, xo, 0) + p(t, to, 0, u),
where ifj and ^ denote the zero-input response of (6.8a), (6.8b) and (6.22a), (6.22b), respectively, while p and p denote the zero-state response of (6.8a), (6.8b) and (6.22a), (6.22b), respectively. The relations (6.29) and (6.30) give rise to the following concepts: Two state-space representations are zero-state equivalent if they give rise to the same impulse response (the same external description). Also, two state-space representations are zero-input equivalent if for any initial state vector for one representation there exists an initial state vector for the second representation such that the zero-input responses for the two representations are identical. The following result is now clear: if two state-space representations are equivalent, then they are both zero-state and zero-input equivalent. They are clearly zero-state equivalent since H{t, r) ^ H{t, r). Also, in view of (6.29) and (6.30), we have Ce^^'-'^^xo = (CP-^)[Pe^^'-'^^p-^]xo = Ce'^^'-'^ho, where (6.26) was used. Therefore, the two state representations are also zero-input equivalent. The converse to the above result is in general not true, since there are representations that are both zero-state and zero-input equivalent, yet not equivalent. In Chapter 5, which deals with state-space realizations of transfer functions, we will consider this topic further. EXAMPLE 6.7. System (6.8a), (6.8b) with 1 0 0 B= -2 - 3 . .1.
C = [-1,-5],
CHAPTER 2 :
Response of Linear Systems
weobtainfrom (6.23) to (6.25) that C^^~^^-^)B+DS(r-T) = CP-^Pe^^'-'^p-^PB^DS(t -T) = Ce^^^-'^^B + D8(t - r), which proves, in view of (6.24) and (6.25), that H(t,
171
D=1
172 Linear Systems
has the transfer function H{s)=C{sI-A)-^B
(.-1)^ (5+l)(5 + 2)'
-5^-1 + 1 5^ + 35 + 2
+D
Using the similarity transformation 1 -1
1" -2
-1
2 -1
1" -1
yields the equivalent representation of the system given by PAP-
B = PB-
c = cp-
^[4,9],
and D = D = 1. Note that the columns of P~^, given by [1,-1]^ and [1,-2]^, are eigenvectors of A corresponding to the eigenvalues Ai = —1,^2 = —2 of A, that is, P was chosen to diagonalize A. Notice that A is in companion form so that its characteristic polynomial is given by 5^ + 3^ + 2 = {s-\-1){s-\-2). Notice also that the eigenvectors given above are of the form [1,A;]^,/ = 1,2. The transfer function of the equivalent representation of the system is now given by 1 H(s) = C{sI-A)-^B
+ D = [4,0]
5+1 0
0 1 5 + 2.
+ 1
-55-1 + 1 = H(s). ( 5 + l ) ( 5 + 2) Finally, it is easily verified that e^'
Pe^'p-K
From the above discussion it should be clear that systems [of the form (6.8a), (6.8b)] described by equivalent representations have identical behavior to the outside world, since both their zero-input and zero-state responses are the same. Their states, however, are in general not identical, but are related by the transformation x{t) = Px{t). In the time-invariant case considered above, transformation P preserves the qualitative properties of the equivalent representations of a system, since in particular, the eigenvalues of A and A are identical. Next, we consider time-varying systems, given by x=A{t)x
+ B{t)u
y = C{t)x + D{t)u,
(6.31a) (6.31b)
where all symbols are defined as in (6.1a), (6.1b). Let P G C^{R^R^^^) and assume that P~^ (t) exists for dllt eR and is continuous. Let X = P(t)x.
(6.32)
T\\Qnx = P{t)x + P{t)x= [P{t)+P{t)A{t)]p-^{t)x + P{t)B{t)u=A{t)x + B{t)udind y = C{t)P~^{t)x + D{t)u = C{t)x + D{t)u. These relations motivate the following definition: the system x=A{t)x
+ B{t)u
y = C{t)x + D{t)u,
(6.33a) (6.33b)
where x = P{t)x, P E C^{R, R^^^), and P ^ is assumed to exist and be continuous for all t G R, and where A(t) = [P(t)A(t) + P(t)]p-\tX Bit) = P(t)B(t), C(t) = C{t)P~^{t), bit) = D(t), is said to be equivalent to the system (6.31a), (6.31b). As in the time-invariant case, the relations between the state transition matrices ^{t, to) and ^(t, to) for the systems of equations
and
X = A(t)x
(6.34)
X = A(t)x,
(6.35)
respectively, and the relations between the impulse responses H(t, r) and H(t, r) of (6.31a), (6.31b) and (6.33a), (6.33b), respectively, are easily established. Indeed, since the solutions of (6.34) and (6.35) are given by (pit, to, xo) = ^it, to)xo and 4>it, to, Xo) = ^(t, to)xo, respectively, we have in view of (6.32) (assuming that P~^ exists for all t G R), P~\t)4>it, to, xo) = ^it, to)[P~\to)xo] or 4>it, to, xo) = Pit)^it, to)P~^ito)xo. Since the solutions of (6.34) and (6.35) are unique, we have that 0(r,T) =
Pit)<^it,T)p-\T)
for all t,T ^R, Recalling that the columns of a fundamental matrix "^ of (6.34) and a fundamental matrix # of (6.35) are linearly independent, it is not hard to show, using (6.32), that # ( 0 - P ( 0 ^ ( 0 for all t G R, Next, recalling that the impulse responses of the equivalent systems (6.31a), (6.31b) and (6.33a), (6.33b) are given by Hit, T) = and
Cit)^it, T)BiT) + Dit)8it - T ) , 0
Hit, T) ^ J Cit)^it, T)BiT) + Dit)8it -T) 0
t^ T, r < T, t^
T, t
respectively, it is easily shown that Hit,
T)
= Hit,
T).
Indeed, we have that Hit, T) = Cit)^it, T)BiT) + Dit)8it - T) = Cit)p-\t)Pit)^it, T)p-\T)PiT)BiT) = Cit)^it, T)BiT) + Dit)8it - T) = Hit, T)
+ Dit)8it - T)
for t > T. We conclude by noting that the notions of zero-state equivalence and zero-input equivalence introduced for time-invariant systems of the form (6.8a), (6.8b) carry over without changes for time-varying systems of the form (6.31a), (6.31b). Furthermore, identically to the time-invariant case, it can be shown that in the case of time-varying systems, if two state representations [such as (6.31a), (6.31b) and (6.33a), (6.33b)] are equivalent, then they are both zero-state and zero-input equivalent. The converse to this statement, however, is not true.
173 CHAPTER 2 :
Response of Linear Systems
174 Lhi^Systems
2.7 STATE EQUATION AND INPUT-OUTPUT DESCRIPTION OF DISCRETE-TIME SYSTEMS In this section, which consists of five subsections, we address the state equation and input-output description of Hnear discrete-time systems. In the first subsection we study the response of Hnear time-varying systems and linear time-invariant systems described by the difference equations (1.15.3a), (1.15.3b) [or (1.1.7a), (1.1.7b)] and (1.15.4a), (1.15.4b) [or (1.1.8a), (1.1.8b)], respectively. In the second subsection we consider transfer functions for linear time-invariant systems, while in the third subsection we address the equivalence of the internal representations of timevarying and time-invariant linear discrete-time systems [described by (1.15.3a), (1.15.3b) and (1.15.4a), (1.15.4b), respectively]. Some of the most important classes of discrete-time systems include linear sampled-data systems that we develop in the fourth subsection. In the final part of this section, we address modes and asymptotic behavior of linear time-invariant discrete-time systems.
A. Response of Linear Discrete-Time Systems We now return to Section 1.15 to consider once again systems described by linear time-varying equations of the form x(k +1) = A(k)x(k) + B(k)u(k) y(k) = C(k)x(k) + D{k)u{k),
(7.1a) (7.1b)
where A:Z^ /?^><^, B : Z-^ T^^^^, C : Z-^ RP""^, md D : Z ^ RP"""^. When A(k) = A,B(k) = B,C(k) = C, and/)(/:) = Z), we have systems described by linear time-invariant equations given by x(k + 1) = Ax(k) + Bu(k) y(k) = Cx(k) + Du(k).
(7.2a) (7.2b)
We recall that in (7.1a), (7.1b) and in (7.2a), (7.2b), x denotes the state vector, u denotes the system input, and y denotes the system output. For given initial conditions ko E Z, x(ko) = Xk^ G R^ and for a given input w, both equations (7.1a) and (7.2a) possess unique solutions x(k) that are defined for all k > ^o^ and thus, the response y(k) for (7.1b) and for (7.2b) is also defined for all k > ko. Associated with (7.1a) is the linear homogeneous system of equations given by xik + 1) = A(k)xik).
(7.3)
We recall from Section 1.15 that the solution of the initial-value problem x(k+l)
= A(k)x(k),
x(ko) = Xk,
(7.4)
is given by x(k) = ^{k, ko)xk, = n
Mj)xk,,
k > ko,
(7.5)
175
where ^(k, ko) denotes the state transition matrix of (7.3) with ^(h
k) = I
(7.6)
[refer to (15.9) to (15.12) in Chapter 1]. Common properties of the state transition matrix (&(/:, /), such as for example the semigroup property given by ^{k, /) = 0 ( ^ , m)<^(m, I),
k>
m>
I,
can quite easily be derived from (7.5), (7.6). We caution the reader, however, that not all the properties of the state transition matrix $(f, r ) for continuous-time systems X = A(t)x carry over to the discrete-time case (7.3). In particular we recall that if for the continuous-time case we have t > r, then future values of the state (p at time t can be obtained from past values of the state cf) at time r, and vice versa, from the relationships (/)(0 = ^(t, T)(f){T) and (/)(T) =^ ^~^{t, r^t) = 0 ( T , 0>(0. i-e., for continuous-time systems a principle of time reversibility exists. This principle is in general not true for system (7.3), unless A~^(k) exists for all k G Z. The reason for this lies in the fact that (k, I) will not be nonsingular if A(k) is not nonsingular for alU. Associated with (7.2a) is the linear, autonomous, homogeneous system of equations given by x(k^
1) = Axik).
(7.7)
From (7.5) it follows that the unique solutions of the initial-value problem x(k + 1) = Ax(k\
x(ko) = Xk^
(7.8)
are given by xik) = ^{h
ko)xk, = A^'-^'hk^,
EXAMPLE 7.1. In (7.3), we let ^o = 0, and A(k) A(k-
1)-"A(0) =
\(k - 1)
(k-
1)2 + 1
a)'-'
0
k ^ ko. Ik 0
(k^ + 1)1
a)'
(7.9) Then ^(k, 0)
If, for example, k = 3, then
"0 [2 5 x(0) x(0). Given 0 0 x(0), we can now readily determine x(3). In view of the form A(0), we have for any x(3) = A(2) • A(l) • A(0) • x(0)
^ > 0, ^{h 0)
To x l
, that is to say, the first column of 0(^, 0) will always be 0 zero for any k. Clearly then, for any initial condition x(0) = a E /?, we have \0] x(k) for all k>0. 0 x(0) = 0
a E: R. The initial state x(0) at
ko ^ 0 for any a G R will map into the state x(l)
. Accordingly, in this case time
EXAMPLE 7.2. In (7.5), let A =
1 0
reversibility will not apply. EXAMPLE 7.3. In (7.8), let A = 1 0
1
1 0
2 . In view of (7.9) we have that A^^ ^o^ = 1
k > ^0, i.e., A^^ ^o) = A when (k - ko) is odd, and
CHAPTER 2 :
Response of Linear Systems
176
j^(k kQ) ^ /when(^-/:o)iseven. Therefore, given ^0 = Oandx(O)
thQnx(k) =
Linear Systems Ax(0)
,k=
1,3,5, ...,andx(y^) - /x(0)
, ^ = 2, 4, 6, . . . . A plot of the
states x(k) = [xi(k), X2(k)]^ is given in Fig. 2.3. iW|
x^ik)
I
2.OH ^
f
T
T
f ^-^
1.0-
1.0
— • - - • - —•- - • — 0
1 2 3 4 5 6 7 8 —
k
0 1 2
3 4 5
6 7 8 '
FIGURE 2.3 Plots of states for Example 7.3 Continuing, we recall that the solutions of initial-value problems determined by linear nonhomogeneous systems (15.13) of Chapter 1 are given by expression (15.14). Utilizing (15.13), the solution of (7.1a) for given x(ko) and u(k) is given as x(k) = ^(k, ko)x(ko) + X
^(^> J + ^)B(jMjX
k > ko,
(7.10)
This expression in turn can be used to determine [as in (15.17) and (15.18) of Chapter 1] the system response for system (7.1a), (7.1b) as y(k) = C(kmk,
ko)x(ko)
k-\
+ ^
C(k)<^(k j + l)B(j)u(j)
+ D{k)u{k),
k > ko,
y(ko) = C(ko)x(ko) + D{ko)u(ko).
(7.11)
Furthermore, for the time-invariant case (7.2a), (7.2b), we have for the system response the expression k-i
y(k) = CA^^'-^'hiko) + ^
CA^-^J-'^^Bu(j) + Du(k),
k > k^, (7.12)
y(ko) = Cx(ko) + Du(ko).
Since the system (7.2a), (7.2b) is time-invariant, we can let ^Q = 0 without loss of generality to obtain from (7.12) the expression k-i
y(k) = CA^x(0) + ^
CA^-^J^^^Bu(j) + Du(kX
k > 0.
(7.13)
j=o
As in the continuous-time case, the total system response (7.11) may be viewed as consisting of two components, the zero-input response, given by il/(k) = C(k)^(k, kQ)x(kol
k > ko,
177
and the zero-state response, given by k-\
CHAPTER 2 :
p(k) = X C(kmk, j + l)B(j)u(j) + D(kMkl p(ko)
k> k0. k = ko
D(ko)u(koX
(7.14)
Finally, in view of (16.20) of Chapter 1, we recall that the (discrete-time) unit impulse response matrix of system (7.1a), (7.1b) is given by H(kJ)
r C(k)^(k, I + l)B(ll = loikl [0,
k>U k=l k
(7.15)
and the unit impulse response matrix of system (7.2a), (7.2b) is given by H{kJ)^lD, [0,
k= h k
(7.16)
and in particular, when / = 0 (i.e., when the pulse is applied at time / = 0), \ CAk- 'B,
Hik,0) = { D,
U
k>0, k=^0. k<0.
(7.17)
EXAMPLE 7.4. In (7.2a), (7.2b), let
0 1" A = 0- 1 .
"11 D = 0. oj We first determine A^ by using the Cay ley-Hamilton Theorem (Theorem 3.1 of Chapter 2). To this end we compute the eigenvalues of A as Ai = 0, A2 ^ - 1 , we let A^ = /(A), where f(s) = s^, and we let g(s) = ais + ao. Then /(Ai) == ^(Ai), or ao = 0 and /(A2) = gi^i), or (-1)^ = -ai + ao. Therefore, A^ = aiA + aol = 8(k) (-l)'-'p(k-l) 0 ,k= 1,2, ...,or A^ -(-1) (-l)'p(k) 0 0 (-1)' , where A^ = /, and where p(k) denotes the unit step given by 0, 1,2, B=
p(k)
0 1. '
C^ =
1, k^O, 0, i^ < 0.
The above expression for A^ is now substituted into (7.12) to determine the response y(k) for /: > 0 for a given initial condition x(0) and a given input u(k), ^ > 0. To determine the unit impulse response, we note that H(k, 0) = 0 for ^ < 0 and k = 0. When k>0, H(h 0) = CA'^-^B = ( - l)^-^p(k - 2) for i^ > 0 or H(K 0) = 0 for i^ = 1 and H(k 0) = (-1)^-^ for y^ = 2, 3, .... • B. The Transfer Function and the z-Transform We assume that the reader is familiar with the concept and properties of the one-sided z-transform of a real-valued sequence {f(k)}, given by
t{f{k)} = f{z) =
J^z-JfU).
(7.18)
Response of Linear Systems
178 Linear Systems
An important property of this transform, useful in solving difference equations, is given by the relation
nf(k + i)} = ^z-^fu + i) = ^zO'-i) / ( ; ) .j=o
= zmf(k)}
- /(O)] = zf(z) - zf(0),
(7.19)
If we take the z-transform of both sides of Eq. (7.2a), we obtain, in view of (7.19), zx(z) - zx(0) = Ax(z) + Bu(z) or x(z) = (zl - AyhxiO)
+ (zl - Ay^Buizy
(7.20)
Next, by taking the z-transform of both sides of Eq. (7.2b), and by substituting (7.20) into the resulting expression, we obtain y(z) = C(zl - Arhx(0)
+ [C(zI - Ay^B
+ D]u{z).
(7.21)
The time sequence {y{k)} can be recovered from its one-sided z-transform y{z) by applying the inverse z-transform, denoted by 2E~^[};(z)]. In Table 7.1 we provide the one-sided z-transforms of some of the commonly used sequences, and in Table 7.2 we enumerate some of the more frequently encountered properties of the one-sided z-transform. The transferfunction matrix H(z) of system (7.2a), (7.2b) relates the z-transform of the output y to the z-transform of the input u under the assumption that ;IL:(0) = 0. We have (7.22)
y(z) = H(z)u(zl where
H(z) = C(zl - Ay^B
(7.23)
+ D.
To relate H(z) to the impulse response matrix H(k, /), we notice that ^{d(k /)} = z~K where 8 denotes the discrete-time impulse (or unit pulse or unit sample) defined in (16.5) of Chapter 1, i.e., 8{k - I)
1, 0,
k = U k¥^L
(7.24)
TABLE 7.1
Some commonly used z-transforms {f(k)},k^^
f(z) = ^{f(k)}
8(k)
1
Pik)
e
1/(1-z-0 z-V(l - z-')^ [z-'il + z-'Wil-z-'f
a'
1/(1-az-')
(k + 1V [(imxk+iy
1/(1-az-'f 1/(1-az-'y^'
k
'(k + DW / > 1 acosak + bsinak
[a + z~^(bsina - acosa)]/(l --Iz
^ cosa + z ^)
TABLE7.2
179
Some properties of z-transforms
CHAPTER 2 :
Time shift —Advance Time shift —Delay Scahng
f(k+l) fik + l) fik-l) f{k-l)
/>1
z'fiz)-zlUz'-'fii--^) z-'f{z)+f(-l) z-'f{z) + l!i=iZ-'+'f{-i)
l>\
Kz/a)
km Convolution Initial value Final value
Response of Linear Systems
m zf{z)-zm
{/(*)},* > o
-z{d/dz)fiz)
m*
f(zmz)
8(k) lT=om8(k-l) = /(/)with/(fc)=0 k
\im,^„ z'fizY lim,^i(l-z-i)/(z)tt
^ If the limit exists. • ' • t i f ( i --^z- i ^^ )f{z) has no singularities on or outside the unit circle.
This implies that the z-transform of a unit pulse applied at time zero is ^ { 5 (^)} = 1. It is not difficult to see nov^ that {H{k, 0)} = ^~^ [y{z)], v^here y{z) = H{z)u{z) with u{z) = 1. This shows that ^-\H{z)]
= ^-\C{zJ-A)-^B
+ D\={H{k,Qi)},
(7.25)
where the unit impulse response matrix H{k^ 0) is given by (1.11). The above result can also be derived directly by taking the z-transform of {H{k,0)} given in (7.17) (prove this). In particular, notice that the z-transform of {A^-^},k= 1,2,... is (zI-A)-^ since
z • =
-1/
\I
-1
2A2
z -A + z ^A
+ •••)
z-Hl-z-'A) {zI-A)--1
(7.26)
Above, the matrix determined by the expression (1 — A)~^ 1 + A + A ^ + --- was used. It is easily shown that the corresponding series involving A converges. Notice also that ^{A^},k = 0,1,2,... is z{zl — A)~^. This fact can be used to show that the inverse z-transform of (7.21) yields the time response (7.13), as expected. We conclude this subsection with a specific example. EXAMPLE 7.5. In system (7.2a), (7.2b), we let C=[1,0], To verify that ^ z(zi-Ay
^[z{zl — A) ^] = A^, we compute -1 z+1
1 " z(z+l) = 1
zTT .
D = 0.
180 Linear Systems
and
8(k) -'[z{zI-Ar'] U ^ 0
A' =
or
as expected from Example 7.4. Notice that 1 z -'[(zI-A)-'] 0
(-lY-'p(k-l) {-l)'p(k)
1 0' 0 1 0 (-1)^-^ 0 (-1)^
when k = 0, when A: = 1, 2,...,
1 z(z + 1) 1
z+l '' 8(k - l)p(k - 1) 8(k - l)p(k - 1) - (-l)^-^p(k - 1) 0 (-l)^-^p(k-l) "0 0 1 0 for for k = 1, and .0 0. 0 1 and 0 ^ - i [ ( z / - A ) - ih] ^ 0
-(-l)^-i'
for yl = 2, 3 , . . . ,
(-i)^-M
which is equal to A^ /: > 0, delayed by one unit, i.e., it is equal to A^~^, k = 1,2,..., as expected. Next, we consider the system response with x(0) = 0 and u(k) = p(k). We have y(k) = ^-'[y(z)] = ^-'[C(zl - AT'B • u{z)] ^"^
1 (z + l)(z - 1)
= _ r 0,
z- 1
z+ 1
\[{i)'-'-{-iY-']p{k-i) ^ = 0,
r 0, k = 0 = lo, k= 1,3,5,..., [l, k = 2,A,6,.... Note that if jc(0) = 0 and M(^) = 6(/:), then y{k) =
1 z{z + 1) 8{k-\)p{k-\)~{-\f-'p{k-\) y^ = 0,1,
iRl %-'[C{zI-A)-'B]
= ^ I 0,
^
which is the unit impulse response of the system (refer to Example 7.4). C. Equivalence of Internal Representations Equivalent representations of linear discrete-time systems are defined in a manner analogous to the continuous-time case.
For systems (7.1 a), (7. lb), we let ^o denote initial time, we let P(k) denote a real nX n matrix that is nonsingular for all k ^ ko, and we consider the transformation jc(^) = P(k)x(k). Substituting the above into (7.1a), (7.1b) yields the system (7.27a)
x(k + 1) = Aik)x(k) + B{k)u{k)
(7.27b)
y(k) = C(k)x(k) + D{k)u{k), where A{k) = P(k +
\)A(k)p-\k)
B(k) = P(k + l}B{k)
(7.28)
C(k) = C(k)p-^ D{k) = D{k). We say that system (7.27a), (7.27b) is equivalent to system (7.1a), (7.1b) and we call P{k) an equivalence transformation matrix. If (^, /) denotes the state transition matrix of (7.3) and ^{k, I) denotes the state transition matrix of xik + 1) = then
Mk,l)
=
Aik)x(k),
(7.29)
P{k)Mk,l)P~\l),
(7.30)
as can be seen by observing that <J>(^,/) = A{k-\)---A{l) = [P(k)A(k-l)P-\k1)]-••[/>(/ + 1)A(1)P-\1)] = P{k)^{k,l)p-\l). In a similar manner as above, it can also be shown that H{k, I) = H(k, I),
(7.31)
where H(k, I) and H(k, I) denote the unit pulse response matrices of systems (7.1a), (7.1b) and (7.27a), (7.27b), respectively. [The reader should verify (7.31).] Thus, equivalent representations of linear discrete-time system (7.1a), (7.1b) give rise to the same unit pulse response matrix. Furthermore, zero-state equivalent representations and zero-input equivalent representations are defined for system (7.1a), (7.1b) in a similar manner as in the case of linear continuous-time systems. Turning our attention now briefly to time-invariant systems (7.2a), (7.2b), we let P denote a real nonsingular nX n matrix and we define x{k) = Px(k).
(7.32)
Substituting (7.32) into (7.2a), (7.2b) yields the equivalent system representation x(k + 1) = Axik) + Bu(k)
(7.33a)
y(k) = Cx(k) + Du(k), where
A = PAP-
B = PB,
c = cp-
(7.33b) D = D.
(7.34)
We note that the terms in (7.34) are identical to corresponding terms obtained for the case of linear continuous-time systems. We conclude by noting that if H(z) and H(z) denote the transfer functions of the unit impulse response matrices of system (7.2a), (7.2b) and system (7.33a), (7.33b), respectively, then it is easily verified that H(z) =" H(z).
181 CHAPTER 2 :
Response of Linear Systems
182 Linear Systems
D. Sampled-Data Systems Discrete-time dynamical systems arise in a variety of ways in the modeling process. There are systems that are inherently defined only at discrete points in time, and there are representations of continuous-time systems at discrete points in time. Examples of the former include digital computers and devices (e.g., digital filters) where the behavior of interest of a system is adequately described by values of variables at discrete-time instants (and what happens between the discrete instants of time is quite irrelevant to the problem on hand); inventory systems where only the inventory status at the end of each day (or month) is of interest; economic systems, such as banking, where, e.g., interests are calculated and added to savings accounts at discrete time intervals only, and so forth. Examples of the latter include simulations of continuous-time processes by means of digital computers, making use of difference equations that approximate the differential equations describing the process in question; feedback control systems that employ digital controllers and give rise to sampled-data systems (as discussed further in the following); and so forth. In providing a short discussion of sampled-data systems, we make use of the specific class of linear feedback control systems depicted in Fig. 2.4. This system may be viewed as an interconnection of a subsystem S\, called tht plant (the object to be controlled) and a subsystem ^2, called the digital controller.
u{t)
x=A(t)x+
B{t)u
y= C(t)x+
D(t)u
y{t )
\ A/D
D/A i
I
w{l<+^) = F(k)w(k) + G{k) y(k) u ik)
u{k) = H{k)w{k) +
Q(k)y(k)
yik)
FIGURE 2.4 Digital control system The plant is described by the equations X = A(t)x + B(t)u y = C(t)x + D(t)u,
(7.34a) (7.34b)
where all symbols in (7.34a), (7.34b) are defined as in (6.1a), (6.1b) and where we assume that ^ > ^ > 0. The subsystem ^2 accepts the continuous-time signal y{t) as its input and it produces the piecewise continuous-time signal u(t) as its output, where t ^ to. The continuous-time signal y is converted into a discrete-time signal {y(k)}, k> k{)> 0, k, ko E Z, by means of an analog-to-digital (A/D) converter and is processed according to a control algorithm given by the difference equations w(k + 1) = F(k)w(k) + G(k)y(k) u(k) = H(k)w(k) + Q(k)y(k),
(7.35a) (7.35b)
where the w(k), y{k), u{k) are real vectors and the F{k), G{k), H{k), and Q{k) are real matrices with a consistent set of dimensions. Finally, the discrete-time signal {u{k)}, ^ > ^0 — 0. is converted into the continuous-time signal u by means of a digital-to-analog (D/A) converter. To simplify our discussion, we assume in the following that ^0 = t]^^. An (ideal) A/D converter is a device that has as input a continuous-time signal, in our case j , and as output a sequence of real numbers, in our case {y{k)}, k = k{), ko + \,..., determined by the relation y{k) = y(tk).
(7.36)
In other words, the (ideal) A/D converter is a device that samples an input signal, in our case y(t), at times toJi,... producing the corresponding sequence {y(tol y(hl •' •}• A D/A converter is a device that has as input a discrete-time signal, in our case the sequence {u{k)}, k = ki^, k^ + \,..,, and as output a continuous-time signal, in our case w, determined by the relation u{t) = u(k), tk^
t < tk+h
k = ko, ko -\- I,,...
(7.37)
In other words, the D/A converter is a device that keeps its output constant at the last value of the sequence entered. We also call such a device a zero-order hold. The system of Fig. 2.4, as described above, is an example of a sampled-data system, since it involves truly sampled data (i.e., sampled signals), making use of an ideal A/D converter. In practice the digital controller ^2 uses digital signals as variables. In the scalar case, such signals are represented by real-valued sequences whose numbers belong to a subset ofR consisting of a discrete set of points. (In the vector case, the previous statement applies to the components of the vector.) Specifically, in the present case, after the signal y{t) has been sampled, it must be quantized (or digitized) to yield a digital signal, since only such signals are representable in a digital computer. If a computer uses, e.g., 8-bit words, then we can represent 2^ = 256 distinct levels for a variable, which determine the signal quantization. By way of a specific example, if we expect in the representation of a function a signal that varies from 9 to 25 volts, we may choose a 0.1-volt quantization step. Then 2.3 and 2.4 volts are represented by two different numbers (quantization levels); however, 2.315, 2.308, and 2.3 are all represented by the bit combination corresponding to 2.3. Quantization is an approximation, and for short wordlengths may lead to significant errors. Problems associated with quantization effects will not be addressed in this book. In addition to being a sampled-data system, the system represented by Eqs. (7.34) to (7.37) constitutes a hybrid system as well, since it involves descriptions given by ordinary differential equations and ordinary difference equations. The analysis and synthesis of such systems can be simplified appreciably by replacing the description of subsystem Si (the plant) by a set of ordinary difference equations, valid only at discrete points in time t^, k = 0, 1, 2, [In terms of the blocks of Fig. 2.4, this corresponds to considering the plant Si, together with the D/A and A/D devices, to obtain a system with input U(k) and output y(k), as shown in Fig. 2.5.] To accomplish this, we invoke the variation of constants formula in (7.34a) to obtain x(t) = ^(tJkXh)
+ I ^(t,T)B{T)u(T)dT,
(7.38)
183 CHAPTER 2 :
Response of Linear Systems
184 Linear Systems
u(k)^
D/A
^(0
x=A(t)x+
B(t)u
y= C(t)x+
D(t)u
y(t)
A/D
y(k)
FIGURE 2.5 System described by (7.40) and (7.43) where the notation ^{t,tk,x{tk)) = x{t) has been used. Since the input u{t) is the output of the zero-order hold device (the D/A converter), given by (131), we obtain from (7.38) the expression /
:^{tk+i)=^{tk+htk)x{tk)-
0(f^+i,T)5(T)JT u{tk).
(7.39)
Jtj,
Since x{k) = x{tk) and u{k) = u{tk), we obtain a discrete-time version of the state equation for the plant, given by x{k^l)
=A{k)x{k)+B{k)u{k),
(7.40)
where (7.41)
^{tk+i,T)B{T)dT.
Next, we assume that the output of the plant is sampled at instants ?[ that do not necessarily coincide with the instants t^ at which the input to the plant is adjusted, and we assume that h
y{t'k)=c{ti)^it'„tk)x{tk)-
u{tk)+D{t'k)u{tk).
(7.42)
Jth
Defining y{k) = y{t'j^, we obtain from (7.42), y{k)=C{k)x{k) where
C{k) ^
+ D{k)u{k),
(7.43)
C{t',)^{ify)
D{k)^C{t',)
f'^{t[,x)B{x)dx
+ D{t'^.
(7.44)
Summarizing, (7.40) and (7.43) constitute a state-space representation, valid at discrete points in time, of the plant [given by (7.34a)] and including the A/D and D/A devices [given by (7.36) and (7.37); see Fig. 2.5]. Furthermore, the entire hybrid system of Fig. 2.4, valid at discrete points in time, can now be represented by Eqs. (7.40), (7.43), (7.35a), and (7.35b). We now turn briefly to the case of the time-invariant plant, where A{t) = A^B{t) = 5,C{t) = C, and D{t) = D, and we assume that f/.+i —tk = T and tj^ — tk = a for alU = 0,1,2,.... Then the expressions given in (7.40), (7.41), (7.43), and (7.44) assume the form x{k+l)=Ax{k)^Bu{k) y{k)=Cx{k)+Du{k),
(7.45a) (7.45b)
185 where
A = €"",8
„AT
=
C = Ce^", D
dT\B,
CHAPTER 2 :
-^ij?"-
dr \B + D.
(7.46)
If t[ = tk, or a = 0 , then C = C stnd D = D. In the preceding, T is called the sampling period and 1/T is called the sampling rate. Sampled-data systems are treated in great detail in texts dealing with digital control systems and with digital signal processing. EXAMPLE 7.6. In the control system of Fig. 2.4, let A =
D = 0,
C = [1,0],
B =
let T denote the sampling period, and assume that a = 0. The discrete-time state-space representation of the plant, preceded by a zero-order hold (D/A converter) and followed by a sampler [an (ideal) A/D converter], both sampling at a rate of 1/r, is given by x(k + 1) = Ax(k) + Bu(k\ y(k) = Cx{k\ where A =
„AT
B^[\
-m
e^^dr\B
0
c =^ c
A^ =
T = dr
=
2 T [1,0].
The transfer function (relating y to u) is given by H(z) = C(zl = [1,0]
[1,0]
Ay^B z-1 0 1 iz-l) 0
2 T
1
(z - ly 1
7^
2 T
T^ (z + 1)
2
(z-ir
The transfer function of the continuous-time system (continuous-time description of the plant) is determined to be H(s) = C(sl - A)~^B = IIs^, the double integrator. The behavior of the system between the discrete instants, t,tk — t < tk+i, can be determined by using (7.38), letting x(tk) = x{k) and u(tk) = u{k). • An interesting observation, useful when calculating A and B, is that both can be e x p r e s s e d i n t e r m s o f a s i n g l e s e r i e s . In particular, A = e^^ = / - h r A - h ( r ^ / 2 ! ) A ^ H • • • = / -h TA'i^iTX where ^ ( 7 ) = / + (r/2!)A + {T^IV.)A^ -h • • • = X^-=o(T^ ( j + l ) ! ) A A T h e n J 5 - {\^ e^'dr)B = (Z^J=Q{TJ^^IU + iyW)B = T^(T)BAf ^ ( r ) is determined first, then both A and B can easily be calculated.
Response of Linear Systems
186
EXAMPLE 7.7. In Example 7.6, ^(7) = I + TA
Linear Systems TA^iT) =
1 T . Therefore, A = I + 0 1
\T^I2 1 T and B = 7^(7)5 = , as expected. 0 1
E. Modes and Asymptotic Behavior of Time-Invariant Systems As in the case of continuous-time systems, we study in this subsection the qualitative behavior of the solutions of linear, autonomous, homogeneous ordinary difference equations x{k + 1) = Ax{k)
(7.47)
in terms of the modes of such systems, where A G R^^^ and x{k) G R^ for every k ^Z^. From before, the unique solution of (7.47) satisfying x(0) = XQ is given by (f>{k, 0, xo) = A^jco.
(7.48)
Let A i , . . . , Ao-, denote the a distinct eigenvalues of A, where A/ with / = 1 , . . . , cr, is assumed to be repeated nt times so that Xr= i ^i = n. Then det(zI-A) = f](z-AO"^
(7.49)
i=i
To introduce the modes for (7.47), we first derive the expressions m-i
AioXfp{k) A^ =/ =X 1
+ 2 Auk(k - ! ) • • • ( / : - / + l)\t^p(k
- /)
/= i
- X[^^o^'/^(^) + ^nfcAf-V(^ - 1) i=l
+ • • • + Ai^n,-i)k(k -l)'"(k-ni
+ 2)Af-^"^-iV(^ " m + 1)], (7.50)
where
A.7 =
1 1 \im{[(z-Xir(zI-Ar']U(ni-l-l) }. II (Hi- l - / ) ! z - . A ,
(7.51)
In (7.51), ['l^^^ denotes the ^th derivative with respect to z, and in (7.50), p(k) denotes the unit step [i.e., p(k) = 0 forfc< 0 and p(k) = 1 forfc> 0]. Note that if an eigenvalue A^ of A is zero, then (7.50) must be modified. In this case, fii-i
^Aul\8(k-l) (7.52) /=o are the terms in (7.50) corresponding to the zero eigenvalue. To prove (7.50), (7.51), we proceed as in the proof of (4.36), (4.37). We recall that {A^} = '^'^[zizl - A)~^] and we use the partial fraction expansion method to determine the z-transform. In particular, as in the proof of (4.36), (4.37), we can readily verify that (7.53) (= 1 /= 0
where the An are given in (7.51). We now take the inverse z-transform of both sides of (7.53). We first notice that e-1
[z(z-AO-^'+'^] =
r'[z-\i - Xiz-'r^'-^'h = f(k - i)p(k - /) f(k - I) 0
for k> I, otherwise.
Referring to Tables 7.1 and 7.2 we note that f(k)p(k) = ^"^[(1 - XiZ'^y^^-^^^] = [l/n(k + 1) • • • (yfc + /)]Af for Ai # 0 and / ^ 1. Therefore, ^-^[llziz - A/)-^^+i)] = llf{k-l)p(k-l) = k(k-l)'"(k-l + l)Xf-Kl^ l.For/ = 0 , w e h a v e S - i [ ( l A/z"^)~^] = Af. This shows that (7.50) is true when A, # 0. Finally, if A, = 0, we note that ^-^[l\z~^] = l\8(k - /), which implies (7.52). Note that one can derive several alternative but equivalent expressions for (7.50) that correspond to different ways of determining the inverse z-transform of z(zl — A)~^ or of determining A^ via some other methods. In complete analogy with the continuous-time case, we call the terms Aiik(k - I)' "(k - I -\- l)Xf~^ the modes of the system (7.47). There are ni modes corresponding to the eigenvalues A/, / = 0 , . . . , n^ - 1, and the system (7.47) has a total of/I modes. It is particularly interesting to study the matrix A^, k = 0,1,2,... using the Jordan canonical form of A, i.e., / = P~^AP, where the similarity transformation P is constructed by using the generalized eigenvectors of A. We recall once more that / = diag[Ji,..., J(j\=diag [/j], where each rit X Ui block // corresponds to the eigenvalue A/ and where, in turn, Ji = diaglJn,..., ///.] with Jfj being smaller square blocks, the dimensions of which depend on the length of the chains of generahzed eigenvectors corresponding to Ji (refer to Subsections 2.3G and 2.4B). Let Jij denote a typical Jordan canonical form block. We shall investigate the matrix J^j, since A^ - p-^j'^P = P'^ diag[jfj]P.
Let
Xi
1
0
0
A/
'•.
0
where
(7.54)
= A// + A^,-,
Jij = 1 A/
0 0 0
1 0
0
0
Ni =
and where we assume that Jij is at X t matrix. Then (jijr
= (A,-/ + NiY ( ^ - l )^'kf'-Nf vi-2^r2 + = Af/ + kXf-'^Ni + ^'^'',,
+ kXiNf'
+ Nf.
(7.55)
187 CHAPTER 2 :
Response of Linear Systems
n Linear Systems
Now since A^^^ = 0 for k ^ t, a. typical t X t Jordan block Jtj will generate terms that involve only the scalars Af, Af~^ . . . , Xf~^^~^\ Since the largest possible block associated with the eigenvalue A/ is of dimension nt X ut, the expression of A^ in (7.50) should involve at most the terms Af, Xf~\ ..., Af"^''''^^ which it does. The above enables us to prove the following useful fact: given A^R^^^, there exists an integer k> 0 such that A^ = 0
(7.56)
if and only if all the eigenvalues A/ of A are at the origin. Furthermore, the smallest k for which (7.56) holds is equal to the dimension of the largest block Jij of the Jordan canonical form of A. The second part of the above assertion follows readily from (7.55). We ask the reader to prove the first part of the assertion. We conclude by observing that when all n eigenvalues A/ of A are distinct, then n
where
A* = ^ A / A f , y t > 0, i=\
(7.57)
A,- = lim[(z - A / ) ( z / - A r M -
(7.58)
If \i = 0, we use 8(k), the unit pulse, in place of Af in (7.57). This result is straightforward, in view of (7.50), (7.51).
0 n
EXAMPLE 7.8. In (7.47) we let A =
The eigenvalues of A are Ai -'4 1 and therefore, rii = 2 and a = I. Applying (7.50), (7.51), we obtain
1
Aio\'lp(k) + 1 0
0 1
AnkX\-'p(k-l) r_ 1 1 P(k) + (k) i 1
_
4
k-i
p(k - 1).
2
-1 2 EXAMPLE 7.9. In (7.47) we let A = . The eigenvalues of A are A1 = - 1 , 0 1 A2 = 1 (so that a = 2). Applying (7.57), (7.58), we obtain ik
^ AioAj + A20A2k _
1 0
0 -1 (-1)' + 0 0
1 1
k^O.
Note that this same result was obtained by an entirely different method in Example 7.3. [0 1 . The eigenvalues of A are Ai [0 - 1 0, A2 = - 1 and o- = 2. Applying (7.57), (7.58), we obtain 1 1 1 z+1 1 Ai = l i m [ z ( z / - A ) - i ] = ,^ 0 0 0 z EXAMPLE 7.10. In (7.47) we let A =
z=0
and
1 0 -1 z+ 1 1 A2 = lim ( z + D 0 z z(z + 1) 0 1 1 1 0 -1" {-l)\k^ 8{k) + A^ = Ai8(k) + A2(-lf = 0 1 0 0
0.
As in the case of continuous-time systems described by (L), various notions 189 of stability of an equilibrium for discrete-time systems described by linear, auCHAPTER 2 : tonomous, homogeneous ordinary difference equations (7.47) will be studied in deResponse of tail in Chapter 6. If 4>(k, 0, Xe) denotes the solution of system (7.47) with x(0) = Xe, Linear Systems then Xe is said to be an equilibrium of (7.47) if (pik, 0, Xe) = Xe for all k> 0. Clearly, Xe = 0 is an equilibrium of (7.47). In discussing the qualitative properties, it is customary to speak, somewhat informally, of the stability properties of (7.47), rather than the stability properties of the equilibrium x^ = 0 of system (7.47). The concepts of stability, asymptotic stability, and instability of system (7.47) are now defined in an identical manner done in Subsection 2.4C for system (L), except that in this case continuous time t{t E 7?+) is replaced by discrete time k{k G Z+). By inspecting the modes of system (7.47) [given by (7.50) and (7.51)], we can readily establish the following stability criteria: 1. The system (7.47) is asymptotically stable if and only if all eigenvalues of A are within the unit circle of the complex plane (i.e., |Ay| < 1, j = 1 , . . . , a). 2. The system (7.47) is stable if and only if |Aj| < 1, j = 1 , . . . , cr, and for all eigenvalues with \Xj\ = 1 having multiplicity rij > 1, it is true that lim [[z - XjTKzI - Ar^]^^J-^-^^] = 0
for/ = l,..,,nj-
1.
(7.59)
z-^Aj
3. The system (7.47) is unstable if and only if (2) is not true. EXAMPLE?.11. The system given in Example 7.8 is asymptotically stable. The system given in Example 7.9 is stable. In particular, note that the solution (j){k, 0, x(0)) = A^x(O) for Example 7.9 is bounded. • When the eigenvalues A/ of A are distinct, then as in the continuous-time case [refer to (4.42), (4.43)] we can readily show that
A'=Y.^j4^J='^J^J^
^-0,
(7.60)
where the vy and Vj are right and left eigenvectors of A corresponding to Ay, respectively. If Ay = 0, we use 8{k), the unit pulse, in place of A^ in (7.60). In proving (7.60), we use the same approach as in the proof of (4.42), (4.43). We have A^ = Qdiag [\\,..., A^]g~^ where the columns of Q are the n right eigenvectors and the rows of Q~^ are the n left eigenvectors of A. As in the continuous-time case [system (L)], the initial condition x(0) for system (7.47) can be selected to be colinear with the eigenvector vt to eliminate from the solution of (7.47) all modes except the ones involving Af. EXAMPLE 7.12. As in Example 7.9, we let A =
-1
21 . Corresponding to the eigen-
0 ij
1, we have the right and left eigenvectors vi = (1, 0)^, V2 values Ai = 1,A2 (1,1)^ VI = (1, 1), andv2 = (0,1). Then V I V I A [ + V2V2A2
"1 - 1 0 I (-1)^ + (1)*, 0. 0 1. 0
yfc> 0.
190 Linear Systems
Choose x(0) = a(l, 0)^ - avi with a 9^ 0. Then
which contains only the mode associated with Ai = - 1 . We conclude the discussion of modes and asymptotic behavior by briefly considering the state equation x(k + 1) = Ax(k) + Bu(k),
(7.61)
where x, u, A, and B are as defined in (7.2a). Taking the 2X-transform of both sides of (7.61) and rearranging yields x{z) = z(zl - Ar^x(0)
+ (zl - Ar^Bu(z).
(7.62)
By taking the inverse 2-transform of (7.62), we see that the solution cf) of (7.61) is the sum of modes that correspond to the singularities or poles of z(zl - A)~^x(0) and of (z/ - A)~^Bu{z). If in particular, system (7.47) is asymptotically stable [i.e., for x{k + 1) = Ax{k), all eigenvalues A^ of A are such that \Xj\ < \, j = \,.. .,n\ and if u{k) in (7.61) is bounded [i.e., there is an M such that \ui{k)\ < M for all k> 0,i = I,..., m], then it is easily seen that the solutions of (7.61) are bounded as well.
2.8 AN IMPORTANT COMMENT ON NOTATION For the most part Chapters 1 and 2 are concerned with the basic (qualitative) properties of systems of first-order ordinary differential equations, such as, e.g., the system of equations given by X = Ax,
(8.1)
where x E. R^ and A G fi^xn^ ^^ ^^^ arguments and proofs to establish various properties for such systems, we highlighted the solutions by using the (/)-notation. Thus, the unique solution of (8.1) for a given set of initial data (^Q, XQ) was written as ^{t, to, XQ) with )(^, to, xo) = XQ. A similar notation was used in the case of the equation given by X = fit, X)
(8.2)
and the equations given by X = A(t)x + B(t)u
(8.3a)
y = C(t)x + D(t)u,
(8.3b)
where in (8.2) and in (8.3a), (8.3b) all symbols are defined as in (E) (see Chapter 1) and as in (6.1a), (6.1b) of this chapter, respectively. In the study of control systems such as system (8.3a), (8.3b), the center of attention is usually the control input u and the resulting evolution of the system state in the state-space and the system output. In the development of control systems theory, the x-notation has been adopted to express the solutions of systems. Thus, the solution
of (8.3a) is denoted by x(t) [or x(t, to, XQ) when to and XQ are to be emphasized] and the evolution of the system output y in (8.3b) is denoted by y(t). In all subsequent chapters, except Chapter 6, we will also follow this practice, employing the usual notation utilized in the control systems literature. In Chapter 6, which is concerned with the stability properties of systems, we will use the (/)-notation when studying the Lyapunov stability of an equilibrium [such as system (8.1)] and the x-notation when investigating the input-output properties of control systems [such as system (8.3a), (8.3b)].
2.9 SUMMARY In this chapter the response of linear systems to specific inputs (subject to particular initial conditions) was studied in detail. State-space descriptions, as well as impulse (resp., unit pulse) response descriptions and transfer functions were used. Continuous-time, time-varying, and time-invariant systems characterized by statespace descriptions were studied first. The time-invariant case was covered in a separate section (Section 2.4) to provide flexibility in the coverage of the material. Similarly, discrete-time systems were treated in a separate section (Section 2.7). Background material on linear algebra for the present and subsequent chapters was presented in Section 2.2. In greater detail, the solutions of the homogeneous state equation x = A(t)x were characterized first, using fundamental matrices and the state transition matrix 0(/, to) in Section 2.3. The solutions of the nonhomogeneous state equations x = A(t)x + B(t)u were derived in the same section. For time-varying systems, the state transition matrix ^(t, to) can be determined in closed form only in special cases. One such case pertains to time-invariant systems X = Ax, where 0(r, ^o) = ^^(^-^o) Methods of determining the matrix exponential e^^ were addressed in Section 2.4. In addition, the asymptotic behavior and the stability of an equilibrium of linear time-invariant systems x = Ax (in terms of modes and eigenvalues) were also addressed in Section 2.4. Linear periodic systems X = A(t)x, A{t) = A{t + T),t ^ R, were treated in Section 2.5. Impulse response representations (resp., transfer function representations) of linear systems, in terms of state equation and output equation parameters were discussed in Section 2.6. In addition, equivalence of state-space representations were treated in Section 2.6. Discrete-time systems represented by state-space descriptions and by the unit pulse response descriptions we addressed in Section 2.7. Results analogous to the continuous-time case were derived. Discrete-time systems arise frequently in the description of sampled-data systems. Such systems were briefly treated in Subsection 2.7D.
2.10 NOTES As mentioned earlier in Chapter 1, standard references on linear algebra and matrix theory include Birkhoff and McLane [2], Halmos [7], and Gantmacher [6]. For more
191 CHAPTER 2 :
Response of Linear Systems
192 Linear Systems
recent texts on this subject, refer to Strang [16] and Michel and Herget [10]. Our presentation in Section 2.2 is in the spirit of the coverage given in [10]. Our treatment of basic aspects of linear ordinary differential equations in Sections 2.3, 2.4, and 2.5 follows along lines similar to the development of this subject given in Miller and Michel [11]. State-space and input-output representations of continuous-time systems and discrete-time systems, addressed in Sections 2.6 and 2.7, respectively, are covered in a variety of textbooks, including Kailath [9], Chen [4], Brockett [3], DeCarlo [5], Rugh [14], and others. For further material on sampled-data systems, refer to Astrom and Wittenmark [1] and to the early works on this subject that include Jury [8] and Ragazzini and Franklin [12]. Detailed treatments of the Laplace transform and the z-transform, discussed briefly in Sections 2.4 and 2.7, respectively, can be found in numerous texts on signals and linear systems, control systems, and signal processing. The state representation of systems received wide acceptance in systems theory beginning in the late 1950s. This was primarily due to the work of R. E. Kalman and others in filtering theory and quadratic control theory and to the work of applied mathematicians concerned with the stability theory of dynamical systems. For comments and extensive references on some of the early contributions in these areas, refer to Kailath [9] and Sontag [15]. Of course, differential equations have been used to describe the dynamical behavior of artificial systems for many years. For example, in 1868 J. C. Maxwell presented a complete treatment of the behavior of devices that regulate the steam pressure in steam engines called flyball governors (Watt governors) to explain certain phenomena. The use of state-space representations in the systems and control area opened the way for the systematic study of systems with multi-inputs and multi-outputs. Since the 1960s an alternative description is also being used to characterize time-invariant MIMO control systems that involves usage of polynomial matrices or differential operators. Some of the original references on this approach include Rosenbrock [13] and Wolovich [17]. This method, which corresponds to system descriptions by means of higher order ordinary differential equations (rather than systems of first-order ordinary differential equations, as is the case in the state-space description) is addressed in Chapter 7.
2.11 REFERENCES 1. K. J. Astrom and B. Wittenmark, Computer-Controlled Systems. Theory and Design, Prentice-Hall, Englewood Cliffs, NJ, 1990. 2. G. Birkhoff and S. MacLane, A Survey of Modern Algebra, Macmillan, New York, 1965. 3. R. W. Brockett, Finite Dimensional Linear Systems, Wiley, New York, 1970. 4. C. T. Chen, Linear System Theory and Design, Holt, Rinehart and Winston, New York, 1984. 5. R. A. DeCarlo, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1989. 6. F. R. Gantmacher, Theory of Matrices, Vols. I, II, Chelsea, New York, 1959. 7. P. R. Halmos, Finite Dimensional Vector Spaces, Van Nostrand, Princeton, NJ, 1958. 8. E. I. Jury, Sampled-Data Control Systems, Wiley, New York, 1958. 9. T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980.
10. A. N. Michel and C. J. Herget, Applied Algebra and Functional Analysis, Dover, New York, 1993. 11. R. K. Miller and A. N. Michel, Ordinary Differential Equations, Academic Press, New York, 1982. 12. J. R. Ragazzini and G. F. Franklin, Sampled-Data Control Systems, McGraw-Hill, New York, 1958. 13. H. H. Rosenbrock, State Space and Multivariable Theory, Wiley, New York, 1970. 14. W. J. Rugh, Linear System Theory, Second Edition, Prentice-Hall, Englewood Cliffs, NJ, 1996. 15. E. D. Sontag, Mathematical Control Theory. Deterministic Finite Dimensional Systems, TAM 6, Springer-Verlag, New York, 1990. 16. G. Strang, Linear Algebra and Its Applications, Harcourt, Brace, Jovanovich, San Diego, 1988. 17. W. A. Wolovich, Linear Multivariable Systems, Springer-Verlag, New York, 1974. 18. L. A. Zadeh and C. A. Desoer, Linear System Theory—The State Space Approach, McGraw-Hill, New York, 1963.
2.12 EXERCISES 2.1. (a) Let {V,F) = (R^, R). Determine the representation of v = (1, 4, 0)^ with respect to the basis v^ = (1, - 1 , 0)^, v^ = (1, 0, - 1 ) ^ , and v^ = (0,1, 0)^. (b) Let V = F^ and let F be the field of rational functions. Determine the representation of V = (^s" + 2, 1/5, - 2 ) ^ with respect to the basis {v^ v^, v^} given in (a). 2.2. Find the relationship between the two bases {v\ v^, v^} and {v^ v^, v^} (i.e., find the matrix of {v^ v^, v^} with respect to {v\ v^, v^}), where v^ = (2, 1,0)^, v^ = ( 1 , 0 , - l ) ^ v 3 = ( l , 0 , 0 / , v i = ( l , 0 , 0 f , v 2 ^ ( 0 , 1 , - 1 ) , and v^ = (0,1,1). Determine the representation of the vector ^2 = (0, 1,0)^ with respect to both of the above bases. 2.3. Let a E: R he fixed. Show that the set of all vectors (x, ax)^, x E R, determines a vector space of dimension one over F = R, where vector addition and multiplication of vectors by scalars is defined in the usual manner. Determine a basis for this space. 2.4. Show that the set of all real nX n matrices with the usual operation of matrix addition and the usual operation of multiplication of matrices by scalars constitutes a vector space over the reals [denoted by (/?"^", R)]. Determine the dimension and a basis for this space. Is the above statement still true if /?"^" is replaced by R^^^^ the set of real mX n matrices? Is the above statement still true if i?"^" is replaced by the set of nonsingular matrices? Justify your answers. 2.5. Let v^ = (s'^, s)^ and v^ = (1,1/^)^. Is the set {v\ v^} hnearly independent over the field of rational functions? Is it linearly independent over the field of real numbers? 2.6. Determine the rank of the following matrices, carefully specifying the field: J (a) -1 where j
1 (b) 7
4 0
s+4 ^2-1 (c) 0 s
-2 6 2s+ 3 -s + 4
(d)
s+ 1
193 CHAPTER 2 :
Response of Linear Systems
194 Linear Systems
2.7. Let V and W be vector spaces over the same field F and let ^2/ : V ^ M^ be a linear transformation. Show that if {^v^,..., ^v^} is a linearly independent set, then so is the set { v \ . . . , v"}. Give an example to show that the converse of this statement is not true. 2.8. Let V and W be vector spaces over the same field F and lot ^ :V ^W ho 3. linear transformation. Show that ^2/ is a one-to-one mapping if and only if J^{s^) = {0}. 2.9. Let ^ = [5,A5,...,A"-i5] and C CA CA"" where A G /?"><",5 G Z^^^^"', and C G /?^><". (a) Prove that if 7]^ G ^ ( ^ ) , then AT]^ G ^ ( ^ ) . (7]^ denotes the coordinate representation of a vector v^ G /?" with respect to the natural basis {^1,..., ^„}.) (b) Prove that if 7]i G ^ ( ^ ) , then A7]i G ^ ( ^ ) . The above shows that J^{^) and ^ ( ' ^ ) are invariant vector spaces under a transformation s^ that is represented by the matrix A. 2.10. Show that a c
d -c
b
d
, where /S. = ad — bc^O. a
2.11. Determine the determinant, the (classical) adjoint, and the inverse of the matrix 45+3 3
s2-2 2.12. Determine the matrix X in
A O
"A-i
5 D
X
O
and Z) are nonsingular. Also, determine the matrix
2.13. (a)
Show that J^r
A C
O Z)
1
D
A C
O D
A O
O D
I D-^C
O I
If A is nonsingular, show that det
Hint: Note that
(c)
C
where it is assumed that A
{detA){det D), where A and Z) are square matrices.
Hint: For Z) nonsingular, use the identity (b)
D-\ A O
A C
B D
A C
B D A O
O I
(detA)det(D-CA-^B). I C
A-^B and D
/ -C
O I
I C
A-^B D
\I A-^B [o D-CA-^B\ In part (b), derive an expression for the case when it is known only that D is nonsingular.
2.14. Show that ^(^1+^2)^ = e^i^e^^t if A1A2 = A2A1.
2.15. Determine the characteristic and the minimal polynomials of the matrices 1 1 0 0 0 1 1 0 0 0 1 0 0 0 0 1
1 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1
195
1 1 0 0 0 1 0 0 A3 = 0 0 1 1 0 0 0 1
CHAPTER 2 :
A4 = h
Hint: These matrices are in Jordan canonical form. 2.16. Determine the Jordan canonical form of the "2 0 0' "2 Ax = 1 2 0 A2 = 1 .2 0 2. .0
matrices 0 0" 2 0 1 2_
A3 =
"2 0 .0
0 0" 2 0 1 2_
2.17. Show that there exists a similarity transformation matrix P such that 0 0
1 0
0 1
••
0 0
•• •
-Oin-l
PAP-^ = Ar = 0 -ao
0
1
0 -a2
-ai
^b] isn,
if and only if there exists a vector b G R^ such that the rank of [b, Ab, i.e., p[/7,AZ?, ...,A"-iZ7] = n.
2.18, Show that if A/ is an eigenvalue of the companion matrix Ac given in Exercise 2.17, then a corresponding eigenvector is v' = (1, A/,..., Af~^)^. 2.19. Let A/ be an eigenvalue of a matrix A and let v' be a corresponding eigenvector. Let /(A) = S l = o ^k^^ be a polynomial with real coefficients. Show that /(A^) is an eigenvalue of the matrix function /(A) = ^[==oOCkA^. Determine an eigenvector corresponding to /(A/). 2.20. For the matrices "1 Ai = 0 .0
2 0 0
0" 2 1.
and
0 0 A2 = 0 0
1 0 0 0 1 0 0 0 1 0 0 0
determine the matrices A}"", A^"", g^l^ and e^^\ t E /? 2.21. Determine some bases for the range and null spaces of the matrices "3 ri 1] and Ai = [1 0 1], 3 Ao = 0 0 _i 0. .3 2.22. Determine all solutions of the equation Arj = v, where
ro A= 1 [2
1 2 0
1 3 2
2 -11 4 - 1 0 2J
and
2.23. Let (j)x{t) = e~^ for ^ G [-1,1] and let 2{t) =
^G[-1,0], t G [0, 1].
2 2 2
r1 1.
Response of Linear Systems
196 Linear Systems
Show that (^1 and (/)2 are Hnearly independent over the field of the real numbers on [-1,1], but not on [0,1]. Remark: This example illustrates the fact that linear independence of time functions over a time interval [a, b] does not necessarily imply linear independence over a time subinterval [a\b'] C [a, b]. 2.24. Show that if two time functions (t)\(t), (f)2{t) are linearly independent over a field F on a time interval [a, b\ then they are linearly independent over F on any interval that contains [a, Z?]. Give a specific example. 2.25. Prove that for A G C[R, /?«><«], (3.14) is true if and only if (3.21) is true for all t,T^R. 2.26. Determine the state transition matrix ^{t, to) for (LH) with [0 01 A(t) = t
0
by (a) directly solving differential equations, (b) using the Peano-Baker series, and (c) using (3.15). 2.27. Determine the state transition matrix ^{t, to) for {LH) with A{t) =
/ 1
0 and detert
mine in this case the solution for (LH) when x(l) = (1,1)^. 2.28. Verify that (/>i(0 = (1/^^ - 1 / 0 ^ and <^2(0 = (2/t\ -llff with 4 2-
are two solutions oi{LH)
A(0 = 0 (a) Determine the state transition matrix ^(t, r) for this system. (b) Determine a solution (j) for this system that satisfies the initial conditions x(l) =
2.29, Given is the system of first-order ordinary differential equations x = f-Ax, where A E ^nxn ^^^ t Ei R. Determine the state transition matrix ^{t, to). Apply your answer to 0 2A ^ the specific case when t^A [2t^ -t\ 2.30. Show that the two linear systems 0 2-t^ and
x(2) =
t 1
1 It 1 x(^> ^ A2{t)x^^^ t
are equivalent state-space representations of the differential equation y - Ity - (2 - i')y = 0. (a) For which choice is it easier to compute the state transition matrix ^{t, toft For this case, compute ^{t, 0). (b) Determine the relation between x^^^ and y and between x^^^ and y. 2.31. Using the Peano-Balcer series, show that when A(0 = A, then (^, ^ ) = e^^^ ^o).
2.32. For {LH) with A{t) =
[ 0
-Ij
determine lim^^oo (f){t, to, XQ) if x(0) = (0,1)^. This
CHAPTER 2 :
example shows that an attempt of trying to extend the concept of eigenvalue from a constant matrix A to a time-varying matrix A{t), for the purpose of characterizing the asymptotic behavior of time-varying systems (LH), will in general not work. 2.33. For the system X = A(t)x + B(t)u,
(11.1)
where all symbols are as defined in (6.1a), derive the variation of constants formula (3.10), using the change of variables z{t) = ^(to, t)x{t). 2.34. For (11.1) with x(fo) = Xo, show under what conditions it is possible to determine M(r) so that (f){t, to, xo) = xo for all t > to. Use your result to find such u{t) for the particular case X = X + e~^u. 2.35. Show that (d/dT)^(t, r) = - 0 ( ^ , T)A(T) for all
t,T^R.
2.36. Determine the state transition matrix ^{t, to) for the system of equations x — e~^^Be^^x, where A E R^^^ and B G R^^^, Investigate the case when in particular AB = BA. 2.37. The adjoint equation of (LH) is given by z
-A(t)^z.
(11.2)
Let 0(r, to) and ^a(t, to) denote the state transition matrices of (LH) and its adjoint equation, respectively. Show that 0^(r, ^o) = [^(to, t)V2.38. Consider the system described by A(t)x + B(t)u
(11.3a)
C(t)x,
(11.3b)
where all symbols are as in (6.1a), (6.1b) with D(t) = 0, and consider the adjoint equation of (11.3a), (11.3b), given by
z = -A(tfz + C(tYv w = B(tfz.
(11.4a) (11.4b)
(a) Let H(t, r) and Ha(t, r) denote the impulse response matrices of (11.3a), (11.3b) and (11.4a), (11.4b), respectively. Show that at the times when the impulse responses are nonzero, they satisfy H(t, r) = Ha(r, tY. (b) If A(0 ^ A, B(t) ^ B, and C(t) ^ C, show that H(s) = -Ha(-sf, where H(s) and Ha(s) are the transfer matrices of (11.3a), (11.3b) and (11.4a), (11.4b), respectively. 2.39. Show that if for (L//), A(t) =
An(t) 0
An(t) A22(0.
where A\i(t), An(t), and A22(t) are submatrices of appropriate dimensions, then ^(t, to)
197
[C|>ll(^,^) 0
Oi2(r,/o)l ^22(tjQ)\
Response of Linear Systems
198 Linear Systems
where ^uit) satisfies the matrix equation {d/dt)^ii(t,to) =Aii(t)^ii(t,to) and where the matrix ^uit^to) satisfies the equation {d/dt)^i2(t,to) =Aii(t) ^n{t,to)+An{t)^22{t,to) with 012(^0,^0) = O. Use the above result to determine the state transition matrix 0(r, 0) for -1 e^n A(t) 0 -ly 2.40. Compute e^^ for
[1
4 2 0
p
[0 2.41. Given is the matrix
10] 0
2]
0" -1 0 -1 0 0 -2 0 (a) Determine e^\ using the different methods covered in this text. Discuss the advantages and disadvantages of these methods. (b) For system (L) let A be as given. Plot the components of the solution (p(t,to,xo) whenXQ = x(0) = (1,1,1)^ and XQ = x(0) = ( | , 1,0)^. Discuss the differences in these plots, if any. 1 2
2.42. Show that for A:
cos bt -sinbt
we have e^^
sinbt cos bt
2.43. Given is the system of equations XI
=
-1 0
0" 1
.•^2.
+
r 1
withx(0) = ( l , 0 ) ^ a n d u{t)=p{t)-
ri,
t >0
\0,
eelsewhere.
Plot the components of the solution of (j). For different initial conditions x(0) {a,bY, investigate the changes in the asymptotic behavior of the solutions. 0 1 is called the harmonic oscillator (refer to Chap-1 0 ter 1) because it has periodic solutions (\>{t) = (0i(^),02(O)^- Simultaneously, for the same values of t, plot (pi (t) along the horizontal axis and (p2(t) along the vertical axis in the X1-X2 plane to obtain a trajectory for this system for the specific initial condition x(0) = x o = (xi(0),X2(0))^ = (1,1)^. In plotting such trajectories, time t is viewed as a parameter, and arrows are used to indicate increasing time. When the horizontal axis corresponds to position and the vertical axis corresponds to velocity, the X1-X2 plane is called the phase plane and 0i,02 (resp. x\,X2) are called phase variables.
2.44. The system (L) with A
2.45. There are various ways of obtaining the coefficients ai{t) given in (2.101). One of these was described in Subsection 2.2J. In the following, we present another method. We consider the relation {d/dt)e^^ = Ae^^ and we use (2.101) to obtain J n—\
dt
n—2
2 a y ( O A ^ " = A 2 a y ( O A ^ " + a „ _ i ( 0 [ - K - i A " " ^ + --- + «iA + ao/)], 7=0
7=0
(11.5)
where the Cay ley-Hamilton Theorem was used. The coefficients ai(t) that satisfy this relation generate a matrix O = T.cCj(t)AJ that satisfies the equation O = AO. For O to equal e^\ we also require that 0(0) = Eaj(0)A^ = / (why?), (a) Show that the ccj{t) can be generated as solutions of the system of equations cco{t) ai{t)
"0 1 0
(b)
0 0
-ao —a\
0
1
cco{t) ai{t) (11.6)
-an-i
with ao(0) = l,aj(0) =OJ> 1. Also, show that the aj(t) generated via (11.6) are linearly independent, Express the solution of the equation X =Ax-\-Bu,
(11.7)
where all symbols are as defined in (6.8a) and x(0) = XQ, in terms of aj(t). Also, show that for x(0) = XQ = 0,0(^,0,0) = 0(r) = JJ'jZQAJBwj{t), where Wj{t)=J^aj{t-T)u{T)dT. 2.46.
First, determine the solution (j) of
XI
"0 1
=
1" 0
XI
withx(O) = (1,1)^. Next,
.•^2.
determine the solution (j) of the above system for x(0) = a ( l , — l ) ^ , a e R,a 7^0, and discuss the properties of the two solutions. 2.47.
In Subsection 2.4C it is shown that when the n eigenvalues Xi of a real nxn matrix A are distinct, then e^^ = ^4=1^1^^'^ where A^ = lims^^.[{s — Xi){sl — A)~^] = ViVi [refer to (4.39), (4.40), and (4.43)], where Vi,Vi are the right and left eigenvectors of A, respectively, corresponding to the eigenvalue A^. Show that (a) E L i ^ « ^ ^' where / denotes the nxn identity matrix, (b) AAi = XiAi, (c) AiA = XiAi, (d) AiAj = dijAi, where 5ij = lif i = j and 5ij = 0 when iy^j.
2.48.
Show that two state-space representations {A,B,C,D} and {A,B,C,D} equivalent if and only if CA^B = CA^B,k = 0 , 1 , 2 , . . . , and Z) = 5 .
2.49.
Find an equivalent time-invariant representation for the system described by the scalar differential equation x = sinltx.
2.50.
Consider the system
are zero-state
x = Ax-\-Bu
(11.7a)
y = Cx,
(11.7b)
where all symbols are defined as in (6.8a), (6.8b) with D = 0. Let 0 3 0 0 (a)
1 0 0 -2
0 0 0 0
0 2 1 0
"0 1 0 0
0" 0 0 1
C = [1,0,1,0].
(11.8)
Find equivalent representations for system (11.7a), (11.7b), (11.8), given by x=Ax + Bu
(11.9a)
y = Cx,
(11.9b)
199 CHAPTER 2 :
Response of Linear Systems
200 Linear Systems
(b) 2.51.
Consider the system (11.7a), (11.7b) with B = 0. (a) Let r-1 10" A = and 0 - 1 0 0 0 2 (b)
2.52.
where x = Px, when A is in (i) the Jordan canonical (or diagonal) form, and (ii) the companion form, Determine the transfer function matrix for this system.
If possible, select x(0) in such a manner so that y(t) = te~\t > 0. Determine conditions under which it is possible to assign y{t),t>0, using only the initial data x(0).
Consider the system given by X2
(a)
(b) 2.53.
+
0
3'= [1,0]
1 7
Determine x(0) so that for u(t) = e~^\y{t) = ke~^\ where ^ is a real constant. Determine k for the present case. Notice that y{t) does not have any transient components. Let u{t) = e^K Determine x(0) that will result in y{t) = ke^K Determine the conditions on a for this to be true. What is k in this case?
Consider the system (11.7a), (11.7b) with 0 3 -1 1 (a)
(b) 2.54.
C = [1,1,1].
0" 1 1 0
,
B =
"0 1
0" 0
0 0
1 0
, c=
For x(0) = [1,1,1,1]^ and for u{t) = [1,1]^, ^ > 0, determine the solution (p(t,0,x(0)) and the output y(t) for this system and plot the components 0,(^O,x(O)),/=l,2,3,4and};,(O,/=l,2. Determine the transfer function matrix H{s) for this system.
Consider the system
x{k+l)=Ax{k)+Bu{k) y{k)=Cx{k),
(11.10a) (11.10b)
where all symbols are defined as in (7.2a), (7.2b) with Z) = 0. Let B
C=[l,l],
and let x(0) = 0 and w(^) = 1, ^ > 0. (a) Determine {y{k)},k > 0, by working in the (i) time domain, and (ii) ztransform domain, using the transfer function //(z). (b) If it is known that when u{k) = 0, then y(0) =y{\) = 1, can x(0) be uniquely determined? If your answer is affirmative, determine x(0). 2.55.
Consider y{z) = H{z)u{z) with transfer function H{z) = l/(z + 0.5). (a) Determine and plot the unit pulse response {h(k)}. (b) Determine and plot the unit step response.
201
(c) If
{i:
u(k)
^=1,2,
CHAPTER 2 :
elsewhere,
Response of Linear Systems
determine {y{k)} for k = 0,1,2,3, and 4 via (i) convolution, and (ii) the z-transform. Plot your answer. (d) For u{k) given in (c), determine y{k) as ^ ^ oo.
2.56. Consider the system (11.1 Oa) with x(0) = XQ and ^ > 0. Determine conditions under which there exists a sequence of inputs so that the state remains at XQ, i.e., so that x{k) = XQ for all ^ > 0. How is this input sequence determined? Apply your method to the specific case B-
XQ
2.57. For system (7.7) with x(0) = XQ and ^ > 0, it is desired to have the state go to the zero state for any initial condition XQ in at most n steps, i.e., we desire that x{k) = 0 for any XQ = x(0) and for all k>n. (a) Derive conditions in terms of the eigenvalues of A under which the above is true. Determine the minimum number of steps under which the above behavior will be true. (b) For part (a), consider the specific cases 0 0 0
1 0 0
0 1 0
0 0 0
A2-
1 0 0
0 0 , 0
A3
0 0 0
0 0 0
0 1 0
Hint: Use the Jordan canonical form for A. Results of this type are important in dead-beat control, where it is desired that a system variable attain some desired value and settle at that value in a finite number of time steps. 2.58. Consider the system representations given by -1 0
X(^+1):
0 x{k) + -2
u(k),
y{k) = [l,l]x{k) + u{k) and
0 -2
X(^+1):
1 x{k) + -3
u(k),
y{k) = [l,0]x{k). Are these representations equivalent? Are they zero-input equivalent? 2.59. For the Jordan block given by [Xi 0
1 A,
0 1
0 0
0 0
••• •••
•••
0" 0
202 Linear Systems
where Jtj G R^^[, show that k{k - 1)
'Af
af-'
0
Af
Uf-i
0
0
Af
0 0
0 0
0 0
/:(fc-l)---a-f +2) ._(,_!)
A*
when ^ > ^ - 1. //m^; Use expression (7.55). 2.60. Consider a continuous-time system described by the transfer function H{s) = 4/(^2 + 25 + 2), i.e., })(5) = H(s)u(s). (a) Assume that the system is at rest and assume a unit step input, i.e., u(t) = 1, t > 0, u(t) = 0,t <0, Determine and plot y(t) for t > 0. (b) Obtain a discrete-time approximation for the above system by following these steps: (i) determine a realization of the form (11.7a), (11.7b) of H(s) (see Exercise 2.61); (ii) assuming a sampler and a zero-order hold with sampling period 7, use (7.46) to obtain a discrete-time system representation x(k +1) = Ax(k) + Bu(k)
(11.11a)
y(k) = Cx(k) + Du(k)
(11.11b)
and determine A, B, and C in terms of T. (c) For the unit step input, u(k) = 1 for /: > 0 and u(k) = 0 for /: < 0, determine and plot y(k), A: > 0, for different values of T, assuming the system is at rest. Compare y(k) with y(t) obtained in part (a). (d) Determine for (11.11 a) and (11.lib) the transfer function H(z) in terms of T. Note that H(z) = C(zl - Ay^B + D. It can be shown that H(z) = (1 - z~^M^-^[H(s)/sl^kT}. Verify this for the given H(s). 2.61. Given a proper rational transfer function matrix H(s), the state-space representation {A, B, C, D} is called a realization of H(s) if H(s) = C(sl - Ay^B + D. Thus, the system (6.8a), (6.8b) is a realization of H(s) if its transfer function matrix is equal to H(s). Realizations of H(s) are studied at length in Chapter 5. When H(s) is scalar, it is straightforward to derive certain realizations, and in the following, we consider one such realization. Given a proper rational scalar transfer function H(s), let D = lim^^oo H(s) and let H,p(s) ^ H{s) - D
+ b\s + bo
bn-lS''
s^ + an-\s^ ^ + '" -\- a\s + ao
a strictly proper rational function, (a) Let
0 0
1 0
0 1
0 -ao
0
0
-ai
-a2
••
0 0
0 0
0
1
0 0 B=
C = [bo
by
bn-l]
"•
-Cln-2
~(^n-l_
0 1
and show that {A,B,C,D}
is indeed a reahzation of H{s). Also, show that {A = is a reahzation of H(^s) as well. These two statespace representations are said to be in controller (companion) form and in observer (companion) form, respectively (refer to Subsection 3.4D). (b) In particular find realizations in controller and observer form for (i) H{s) = \/s^, (ii) H{s) = (ol/{s^ + 2l^(OnS+(ol), and (iii) H{s) = {s+ 1)V(^ - 1)^2.62. Given are the systems 5*1 and 5*2 described by the equations X\=A\X\^B\U\
I
X2=A2X2+B2U2
y\ =C\x\+
\
y2= C2X2 + D2U2
D\ u\
(^2),
where all symbols are defined as in (6.8a), (6.8b) with an appropriate set of dimensions for all matrices and vectors. (a) Determine state-space representations for the following composite systems. (i) Systems connected in tandem or in series: u = u^
y = U2
y2 = y S2
FIGURE 2.6 Two systems connected in series (ii) Systems connected in parallel: u^
*J
S^
yi
rf
s
f>
U2
Sz
y2
1
FIGURE 2.7 Two systems connected in parallel (iii) Systems connected in 3. feedback configuration:
^o^ y2
yi
U2
FIGURE 2.8 Feedback configuration Hint: In each case, use
as the state of the composite system.
(b) If Hi(s) is the transfer function matrix of Si,i = 1 , 2 , determine the transfer function matrix for each of the above composite systems in terms of the Hi{s),i = 1 , 2 .
203 CHAPTER 2: Response of Linear Systems
204 Linear Systems
2.63. Assume that H(s) is a p X m proper rational transfer function matrix. Expand H(s) in a Laurent series about the origin to obtain 00
H(s) = Ho+His'^
+ ••• +HkS~^ + ••• =
^HkS'K k=0
The elements of the sequence {HQ, HI, ..., Hj,,...} are called the Markov parameters of the system. These parameters provide an alternative representation of the transfer function matrix H(s) (why?), and they are useful in Realization Theory (refer to Chapter 5). (a) Show that the impulse response H(t, 0) can be expressed as
H(t,0) == Ho8(t) + Y.Hk
(k-l)\
In the following, we assume that the system in question is described by (6.8a), (6.8b). (b) Show that H(s) = D + C(sl - A)-^B = D +
^[CA^-^B}s~^,
which shows that the elements of the sequence {D, CB, CAB,..., CA^ ^B,...} are the Markov parameters of the system, i.e., HQ = D and H^ = CA^~^B, )^ = 1,2, . . . . (c) Show that H(s) = D+ -^C[Rn-is''-^ a(s)
+ • • • + i?i^ + Ro]B,
where a(s) = s'^ + a„_i5"~^ + • • • + a\s + ao = det (si - A), the characteristic polynomial of A, and Rn-i = /, Rn-2 = ARn-i + a „ - i / = A + a „ - i / , . . . , Ro = A«-i +(2„_iA"-2 + ... -^aiL Hint: WviiQ (si-A)-^ = [l/a(s)][adj (si - A)] = [l/a(s)][Rn-is^-^+ '" + Ris + RQ], and equate the coefficients of equal powers of s in the expression a(s)I = (si - AMn-is""'^
+ '" -hRis-h
Rol
2.64. Given the transfer function of a system, suggest different methods to determine its Markov parameters. Apply these methods to the specific cases given by s 1 1 s+ 1 H(s) = (s^ - l)/(s^ + 2 ^ + 1 ) and H(s) 0 2.65. The frequency response matrix of a system described by its p X m transfer function matrix evaluated ats = jco, H(co) ^
H(s)l=j^,
is a very useful means of characterizing a system, since typically it can be determined experimentally, and since control system specifications are frequently expressed in terms of the frequency responses of transfer functions. When the poles of H(s) have negative real parts, the system turns out to be bounded-input/bounded-output (BIBO) stable (refer to Chapter 6). Under these conditions, the frequency response H(co) has a clear physical meaning, and this fact can be used to determine H(a)) experimentally.
(a) Consider a stable SISO system given by y{s) = H{s)u{s). Show that if u{t) = ksm{(Oot + (j)) with k constant, then y{t) at steady-state (i.e., after all transients have died out) is given by yss{t) = k\H{(Oo)\sm{(Oot + 0 + 0{(Oo)), where \H{Q))\ denotes the magnitude of //(co) and 6{Q)) = arg H{Q)) is the argument or phase of the complex quantity H{(o). From the above it follows that H{(o) completely characterizes the system response at steady-state (of a stable system) to a sinusoidal input. Since u{t) can be expressed in terms of a series of sinusoidal terms via a Fourier series, H{(o) characterizes the steady-state response of a stable system to any bounded input u{t). This physical interpretation does not apply when the system is not stable. (b) For the /? X m transfer function matrix H{s), consider the frequency response matrix H{(o) and extend the discussion of part (a) above to MIMO systems to give a physical interpretation of//(co). 2.66. Let A G /?"><" and B e /?"><'^. (a) Is it true that rank [B,AB,... ^A'^'^B] = rank [5,A5,... ,A"-i5,A"5]? Justify your answer. (b) Determine conditions under which rank [B,AB,.. .,A"^~^B] = rank [A5,..., A"~^5,A"5]. Hint: Use the Sylvester Rank Inequality, which relates the rank of the product of two matrices to the ranks of the individual matrices, rankX + rankY — n< rank(XY) < imn{rankX,
rankY},
where X e /?^><" and Y e /?"><'^. 2.67. (Double integrator) (a) Plot the response of the double integrator of Example 7.6 to a unit step input. (b) Consider the discrete-time state-space representation of the double integrator of Example 7.6 for 7 = 0.5,1,5 sec and plot the unit step responses. (c) Compare your answers in (b) with your result in (a). 2.68. (Economic model for national income) [D. G. Luenberger, Introduction to Dynamic Systems, Wiley, 1979.] A simple model describing the national income dynamics can be formulated in discrete times as follows. The national income y{k) in year k in terms of consumer expenditure c{k), private investment i{k), and government expenditure g{k) is assumed to be given by y{k) = c{k) + i{k) + g{k), where the interrelations between these quantities are specified by c ( ^ + 1) = ccy{k) and / ( ^ + 1) = ^[c{k — 1) — c{k)\. The constant a is called the marginal propensity to consume, while /3 is a growth coefficient. Typically, 0 < a < 1 and /3 > 0. From these assumptions we obtain the difference equations c ( ^ + 1) = ac(^) + ai{k) -\- ag{k),i{k-\-1) = (Pa — P)c{k) -\-Pai{k) + /3ag{k), with discrete-time statespace representation given by XI ( ^ + 1 )
X2(^+l)
a p{a-l)
y{k) = [\A]
a pa
xi{k) X2{k)
a u{k)
pa
XI (k) + u{k), X2{k)
where x\ {k) = c{k),X2{k) = i{k), and u{k) = g{k). Let the parameters a,/3 take the values (i) a = 0.75,/3 = 1 (ii) a = 0.75,/3 = 1.5, and (iii) a = 1.25,p = 1.
205 CHAPTER 2 :
Response of Linear Systems
206 Linear Systems
(a) Determine the eigenvalues of A for all cases and express x(k) when u = Oin terms of the initial conditions and the modes of the system. (b) Plot the states for /: > 0 when u(k) is the unit step and x(0) = [0, 0]^. Comment on your results. (c) Plot the states for/: > Owhenw = Oandx(O) = [5, 1]^. Comment on your results. 2.69. (Spring mass system) Consider the spring mass system of Example 4.1 in Chapter 1. For Ml = 1 kg, M2 = 1 kg, K = 0.091 N/m, Ki = 0 . 1 N/m, K2 = OA N/m, B = 0.0036 N sec/m, Bi = 0.05 N sec/m, and B2 = 0.05 N sec/m the state-space representation of the system in (4.2) of Chapter 1 assumes the form 0 -0.1910 0 0.0910
1 0 -0.0536 0.0910 0 0 0.0036 -0.1910
\xi "0 0 0.0036 \X2 4- 1 1 Us 0 -0.0536 |_X4_ 0
0]
r/r
0 0 [/2.
-ij
A where xi = yi, X2 = yi X3 = y2, and X4 = y2.
(a) Determine the eigenvalues and eigenvectors of the matrix A of the system and express x(t) in terms of the modes and the initial conditions x(0) of the system, assuming that / i = /2 = 0. (b) Forx(O) - [1,0, 0.5, 0]^ and / i = f2 = 0 plot the states for t > 0. (c) Let y = Cx with C =
1 0 0 0 1 0
0]
denote the output of the system. Determine
the transfer function between y and u = [fi, fiV• (d) For zero initial conditions, fi(t) = 8(t) (the unit impulse), and f2(t) = 0, plot the states for r > 0 and comment on your results. (e) It is desirable to explore what happens when the mass ratio M2/M1 takes on different values. For this, let M2 = OLM\ with Mi = 1 kg and a = 0.1, 0.5, 2, 5. All other parameter values remain the same. Repeat (a) to (d) for the different values of a and discuss your results. 2.70. (RLC circuit) For the circuit ofExample 4.2 in Chapter 1, let/?i = 211, i?2 = l i ^ , Ci = 1 mF, C2 = 1 mF, and L - 0.5 H. (a) Determine the eigenvalues of A and express x{i) when v = 0 in terms of the initial conditions and the modes of the system. (b) Plot the states for ? > 0 when v - 0 and x(0) = [5, \, 0]^. Repeat for x(0) = [0, 0, 5]^ and comment on your results. (c) Compute the transfer function between y = [vi, V2, VB]^ and v. (d) Plot the states when the input v is the unit step and x(0) = [0, 0, 0]^. Comment on your results. 2.71. (Armature voltage-controlled dc servomotor) Using a consistent set of units for the armature voltage-controlled dc servomotor in Example 4.3 of Chapter 1, let Ra = 2, La = 0.5, J = I, B = I, KT = 2, and Ke = I. The state-space description of this system is given by (4.8) of Chapter 1, and here assumes the form Xi~
X2 •^3_
"0 1 = 0 -1 .0 - 2
01 \xi' "0 2 \X2 + 0 ea, -4J 1x3. .2
where xi = ^ is the shaft position, X2 = ^ is the angular velocity, X3 = ia is the armature current, and the input w = ^^ is the armature voltage. (a) Determine the eigenvalues and eigenvectors of A and express x{t) in terms of the modes and the initial conditions of the system when Ca = 0.
(b) Plot the states for r > 0 when the input Ca is the unit step, and x(0) Comment on your results.
[0, 0, 0]^
2.72 (Unit mass in an inverse square law force field) Consider Example 11.3 of Chapter 1 where for a satellite, ro - 4.218709065 X 10^ m and COQ = 7.29219108 X 10~^ rad/sec. The linearized model about the orbit d{f) = COQ^ + ^o is given by ' 0
Xi Xl
is Xi\
3col —
0
1 0 0
0 0 0
0"
"0 0 1r \Xi 2ro(0o U2 1 1 + 0
2CL)O
0 r^i 0 1 [^2
Us
X4 0 0 0 0 ^0 ^0-' ^ -• where xi(t) = r{t) - ro, X2{t) = r{t), x^it) = 0(t), and ^4(0 - ^(0 - ^o(a) Determine the eigenvalues of A. Is the system asymptotically stable? Explain your answer. (b) Plot the states for x(0) = [100, 0,0,0]^ and zero input. Comment on your results. (c) Plot the states for wi(0 = 0, U2(t) = -land;\c(0) = 0. If the input represents force imposed on the satellite by friction, comment on your results.
2.73. (Magnetic ball suspension system) Consider the magnetic ball suspension system of Exercise 1.21 in Chapter 1. It can be shown that under certain simplifying assumptions, a linearized model x = Ax + Bu, y = Cxof this system is given by R 0 rli L Ui 0 0 U3 + 0 2Kieq L-^BJ .o_
z
y = [0,1,0] A typical set of parameters is Sgq = 0.01 m, ieq = 0.125 A, M = 0.01058 kg, K = 6.5906 X 10-4 N m^/A^, R = 31.1 n , and L = 0.1097 H. (a) Determine the eigenvalues and a set of eigenvectors of A. (b) Compute the transfer function. (c) Plot the states for ^ > 0 if the ball is slightly higher than the equilibrium position, namely, if x(0) = [0, 0.0025, 0]^. Comment on your results. 2.74. (Automobile suspension system) [M. L. James, G. M. Smith, and J. C. Wolford, Applied Numerical Methods for Digital Computation, Harper & Row, 1985, p. 667.] Consider the spring mass system in Fig. 2.9, which describes part of the suspension system of an automobile. The data for this system are given as mi = \ X (mass of automobile) = 375 kg, m2 = mass of one wheel = 30 kg, ki = spring constant = 1500 N/m, k2 = linear spring constant of tire = 6500 N/m, c = damping constant of dashpot = 0, 375, 750, and 1125 N sec/m, xi = displacement of automobile body from equilibrium position m.
207 CHAPTER 2 :
Response of Linear Systems
208 Linear Systems
,,v 1 . 2nvt u(t)= - s i n —— 6 20
FIGURE 2.9 Model of an automobile suspension system X2 = displacement of wheel from equilibrium position m, V = velocity of car = 9, 18, 27 or 36m/sec. A linear model x = Ax-\-Bu for this system is given by 1
0 kl
C
mi
m\ 0 c
0
h
c — m\ 1
m\ 0 k\ +^2
ni2
_ 1712
0 ]
0
h
r 0 1
^ -, Xl ^2 X3
0 0 h
+
u(t),
X4
Lm2^ 1712]
1712
where u{t) = ^ sin(27rv^/20) describes the profile of the roadway, (a) Determine the eigenvalues of A for all the above cases. (a) Plot the states for r > 0 when the input u{t) = ^ sin(27rvr/20) and x(0) = [0,0,0,0]^ for all the above cases. Comment on your results. 2.75. (Building subjected to an earthquake) [M. L. James, G. M. Smith, and J. C. Wolford. Applied Numerical Methods for Digital Computation, Harper & Row, 1985, p. 686.] A three-story building is modeled by a lumped mass system as shown in Fig. 2.10. For ground acceleration v, the differential equations of motion in terms of mass displacements [^1,^2,^3] relative to the ground are given in state-variable form x = Ax-\-Buhy
ki +^2
1 2c
XI
mi
mi
X2
0
0
h. m2
c
^2 + ^3
2c
m2
m2
m2
0
0
0
0 k2
0 c
0
0
mi
mi
0
0
XI
0
1
0
0
X2
h. m2
c
X\
m2
X5
0
1
X6
X^ X4
is xe
X3
—
0
0
0 k3
ms
0 c ms
k3
c
ms
m3_
+
0 -1 0 - 1 w, 0 -1
where x\ = qi,X2 = qi,X3 = q2,M = q2,xs = q3,x^ = ^3, and u = V. Let k = 3.5025 X lO^N/m, m = 1.0508 x 10^kg, and c = 4.2030 x lO^N sec/m. Investigate
the dynamic response of the structure due to the ground acceleration ufoYT = 0.4,0.6, and 0.8 sec (see Fig. 2.10). In particular: (a) Plot the distortions 1 0 0 0 0 0 XI yi 1 0 0 0 [xi,X2,X3,X4,X5,X6]^. 0 = -1 yi = X3-XI 0 1 0 0 0 -1 X5-X3 y3 If serious damage occurs when a distortion exceeds 0.08 m, will the given ground acceleration due to the earthquake cause serious damage to the building? (b) Repeat (a) for different values of the damping parameter c. In particular, let Qg^ = acoid, where a = 2,3,10, and repeat (a) for each value of a. Also, determine the eigenvalues of A for each a and comment on your results.
/C2 = 2/f
1 gf = 10 m/sec^
4.0
FIGURE 2.10 A model for the dynamics of a three-story structure 2.76. (Aircraft dynamics) [B. Friedland, Control System Design, An Introduction to StateSpace Methods, McGraw-Hill, 1986.] For purposes of control system design, aircraft dynamics are frequently linearized about some operating condition, called 3. flight regime,wh^YQ it is assumed that the aircraft velocity(Mach number)and attitude are constant. The control surfaces and engine thrust are set, or trimmed, to these conditions
209 CHAPTER 2 :
Response of Linear Systems
210 Linear Systems
and the control system is designed to maintain these conditions, i.e., to force perturbations (deviations) from these conditions to zero. It is customary to separate the longitudinal motion from the lateral motion, since in many cases the longitudinal and lateral dynamics are only lightly coupled. As a consequence of this the control system can be designed by considering each channel independently. The aerodynamic variables of interest are summarized in Table 12.1 and Fig. 2.11. The aircraft body axes are denoted by x,y, and z, with the origin fixed at some reference point (typically the center of gravity of the aircraft). The positive directions of these axes are depicted in Fig. 2.11. Roll, pitch, and yaw motions constitute rotations about the x-, y-, and z-axes, respectively, using the following sign convention: looking at Fig. 2.11a we see that the pitch angle 6 increases with upward rotation in the side view shown; in Fig. 2.11c, which gives the top view of the aircraft, yaw angle if/ increases in the counterclockwise direction; and looking at Fig. 2.lid, which provides the front view of the aircraft, we see that roll angle (p increases in the counterclockwise direction. We let co^ = r,(x)y = q, and (o^ = p denote yaw rate, pitch rate, and roll rate, respectively. The velocity vector V is projected onto the body axes with w, v, and w being the projections onto the x-, y- and z-axes, respectively. The angle-of-attack a is the angle that the velocity vector makes with respect to the x-axis in the (positive) pitch direction, and the side-slip angle jS is the angle that it makes with respect to the x-axis in the (positive) yaw direction. Note that for small angles, a — w/u and ^ — vlu. The aircraft pitch motion is typically controlled by a control surface called the elevator, roll is controlled by a pair of ailerons, and yaw is controlled by a rudder. Aircraft longitudinal motion. As a specific example, consider the numerical data for an actual aircraft, the AFTI-16 (a modified version of the F-16fighter)in the landing approach configuration (speed V = 139 mph). The components of the state-space equation x = Ax + Bu that describe the longitudinal motion of the aircraft are given by 32.2] r^i "xA ["-0.0507 -3.861 0 0 i:21 -0.00117 -0.5164 1 -0.0717 0 \x2 + 0 is ~ -0.000129 1.4168 -0.4932 Us - 1 . 6 4 5 u, X4 0 0 1 0 J 1x4 0 where the control input w = 5^ is the elevator angle and the state variables in the vector X = [Aw, a, q, 6]^ are the change in speed, angle of attack, pitch rate, and pitch, respectively. TABLE 12.1
Aerodynamic variables Lateral Rates
p: n j8:
roll rate yaw rate side-slip angle
Positions
0: ij/: x: y:
roll angle yaw angle forward displacement cross-talk displacement
Controls
8A : aileron deflection 8R: rudder deflection
Longitudinal a: q: Aw: 9: z:
8E :
angle of attack pitch rate change in speed pitch angle altitude
elevator deflection
211 CHAPTER 2 :
Response of Linear Systems
(a) Side view
(b) Angle-of-attack a and side-slip angle
Deflected aileron
Deflected rudder
(d) Front View
FIGURE 2.11 Aircraft dynamics The longitudinal modes of the aircraft are called short period and phugoid. The phugoid eigenvalues, which are a pair of complex conjugate eigenvalues close to the imaginary axis, cause the phugoid motion, which is a slow oscillation in altitude, (a) For the state-space model that describes the aircraft longitudinal motion, determine the eigenvalues and eigenvectors of A. Express x{t), when w = 0, in terms of the initial conditions and the modes of the system.
212 Linear Systems
(b) Let the elevator deflection §£ be - 1 for t E [0, T] and zero afterward, where T may be taken to be the sampling period in your simulation. This corresponds to the maneuver made when the pilot pulls back on the stick to raise the nose of the airplane. (The minus sign conventionally represents pulling the stick back.) The elevator must be restored to its original position when the desired new climbing angle is reached or the plane will keep rotating. Plot the states for x(0) = [0, 0, 0, 0]^ and comment on your results. (c) Plot the states for x(0) = [0,0, 0,0]^, using a negative unit step as the elevator input. This happens when the elevator is reset to a new position in the hope of pitching the plane up and climbing. Comment on your results. (d) As a second example, consider the numerical data for a Boeing 747 jumbo jet flying near sea level at a speed of 190 mph. The state-space description x = Ax + Buoi the longitudinal motion is now given by -0.0188 -0.0007 0.000048 0
11.5959 -0.5357 -0.4944 0
0 1 -0.4935 1
32.21 U i 0 0 0 \x2 + -0.5632 0 Us X4 0 0 J
(C. E. Rohrs, J. L. Melsa, and D. G. Schultz, Linear Control Systems, McGrawHill, 1993, p. 92). Repeat (a) to (c) for the present case and discuss your answers in view of the corresponding results for the AFTI-16 fighter. Aircraft lateral motion. As a specific example consider the lateral motion of a fighter aircraft traveling at a certain speed and altitude with state-space description X = Ax + Bu given by Xi X2 X3 X4_
-0.746 -12.9 4,31 0 0.0012 6.05 + -0.416 0
0.006 -0.746 0.024 1
- D.999 0.03690 U i 0 0.387 \x2 -0.174 0 \X3 0 0 J \X4
O.OO92I 0.952 ' -1.76 [U2_ 0
where the control inputs [ui, U2V = [^A,^RV denote the aileron and rudder deflections, respectively, and the state variables in the vector x = [/3, p, r, c^]^ are the side-slip angle, roll rate, yaw rate, and roll angle, respectively. (e) The eigenvalues for the aircraft lateral motion consist typically of two complex conjugate eigenvalues with relatively low damping, and two real eigenvalues. The modes caused by complex eigenvalues are called dutch-roll. One real eigenvalue, relatively far from the origin, defines a mode called roll subsidence, and a real eigenvalue near the origin defines the spiral mode. The spiral mode is sometimes unstable (spiral divergence). Find the modes for the aircraft lateral motion of the fighter. (f) Plot the states when x(0) = [0, 0, 0, 0]^, wi is the unit step and U2 = 0. Repeat for ui = 0 and U2 the unit step. Comment on your results. 2.77. (Read/write head of a hard disk) [MATLAB Control System Toolbox User's Guide, The MathWorks, Inc., 1993.] Using Newton's law, a simple model for the read/write head of a hard disk is described by the differential equation J6 + c0 + k0 = Kj /, where / represents the inertia of the head assembly; c denotes the viscous damping coefficient
of the bearings; k is the return spring constant; Kj is the motor torque constant; 6, 9, and 6 are the angular acceleration, angular velocity, and position of the head, respectively; and / is the input current. A state-space model x = Ax + Bu of this system is given by
0 k J
1] Ui c U2. +
"7-1
0 Kj
-T
where xi ^ e,X2 = 0, and u = i. Let J = 0.01, c = 0.004, k = 10, and KT = 0.05. (a) Determine the eigenvalues of A. With w = 0, is the trivial solution x = 0 asymptotically stable? Explain. (b) Plot the states for r > 0 when the input is the unit step and x(0) = [0, 0]^. (c) Let the plant be preceded by a zero-order hold (D/A converter) and followed by a sampler (an ideal A/D converter), both sampling at a rate of 1/7, where T = 5 ms. Derive the discrete-time state-space representation of the plant. Repeat (a) and (b) for the discrete-time system and comment on your results.
213 CHAPTER 2 :
Response of Linear Systems
CHAPTER 3
Controllability, Observability, and Special Forms
It is frequently desirable to determine an input that causes the states of a system to assume different values in finite time (e.g., to transfer the state vector from one specified vector value to another). Such is the case, for example, in satellite attitude control, where the satellite must change its orientation. This type of desirable property leads naturally to the concepts of state reachability and controllability, which will now be studied at length. Another desirable property of systems is the ability to determine the state from output measurements. Since it is frequently difficult or impossible to measure the state of a system directly (for example, internal temperatures and pressures in an internal combustion engine), it is extremely desirable to determine such states by observing the inputs and outputs of the system over some finite time interval. This leads to the concepts of state observability and constructibility, which will also be studied here. The principal goals of this chapter are to introduce and study in depth the system properties of controllability and observability (and of reachability and constructibility) as well as special forms for the state-space system descriptions when a system is controllable or uncontrollable, and observable or unobservable. These special forms are very useful in the study of the relationships between state-space and input-output descriptions of a system. Note that controllability and observability play a central role when a given impulse response or a transfer function description is realized by means of a state-space description, as will be shown in Chapter 5. These special forms also provide insight into the mechanisms concerning capabilities and limitations of state controllers and state observers, as will be demonstrated in Chapter 4. The concepts of controllability and observability are central in the study of state feedback controllers (resp., output controllers) and state observers. State controllability refers to the ability to manipulate the state by applying appropriate inputs (in particular, by steering the state vector from one vector value to any other vector value in 214
finite time). It turns out that controllability is a necessary and sufficient condition for complete eigenvalue assignment in the system matrix A by means of state feedback. State observability refers to the ability to determine the initial state vector of the system from knowledge of the input and the corresponding output over time. State observability is a necessary and sufficient condition for the arbitrary eigenvalue assignment in an asymptotic state estimator, or state observer that estimates the state of the system using input and output measurements. State feedback controllers and state observers are studied in Chapter 4.
3.1 INTRODUCTION This chapter consists of two parts. In Part 1, consisting of Sections 3.2 and 3.3, the important concepts of state reachability (controllability) and observability (constructibility) are introduced. This is accomplished for continuous- and discrete-time systems that may be time-varying or time-invariant. In Part 2, consisting of Sections 3.4 and 3.5, special forms for state-space representations are developed for controllable or uncontrollable and observable or unobservable time-invariant (continuousand discrete-time) systems. In addition, the Smith-McMillan form of a transfer function matrix and the poles and zeros of a system are introduced and studied. In Subsection A of this section, the concepts of reachability and controllability and observability and constructibility are introduced, using discrete-time timeinvariant systems. In this way, significant insight into the concepts is gained early, together with a clear understanding of what these properties imply for a system. Discrete-time systems are selected for this exposition because the mathematical development is simple in this case, allowing us to concentrate on explaining concepts and their impUcations. The continuous-time case is treated in detail in Sections 3.2 and 3.3.
A. A Brief Introduction to Reachability and Observability Reachability and controllability are introduced first, for the case of discrete-time time-invariant systems, followed by observability and constructibility. Finally, duality is briefly discussed. 1. Reachability and controllability The concepts of state reachability (or controllability-from-the-origin) and controllability (or controllability-to-the-origin) are introduced here and are discussed at length in Section 3.2. In the case of time-invariant systems, a state xi is called reachable if there exists an input that transfers the state of the system x{t) from the zero state to xi in some finite time T. The definition of reachability for the discrete-time case is completely analogous. Figure 3.1 shows that different control inputs ui{t) and U2{t) may force the state of a continuous-time system to reach the value x\ from the origin at different finite times, following different paths. Note that reachability refers to the ability of the
215 CHAPTERS:
Controllability, Observability, and Special Forms
216 Linear Systems
FIGURE 3.1 A reachable state xi
system to reach xi from the origin in some finite time; it specifies neither the time it takes to achieve this nor the trajectory to be followed. A state XQ is called controllable if there exists an input that transfers the state from XQ to the zero state in some finite time T. See Fig. 3.2. The definition of controllability for the discrete-time case is completely analogous. Similar to reachability, controllability refers to the ability of a system to transfer the state from XQ to the zero state in finite time; it too specifies neither the time it takes to achieve the transfer nor the trajectory to be followed. We note that when particular types of trajectories to be followed are of interest, then one seeks particular control inputs that will achieve such transfers. This leads to various control problem formulations, including the Linear Quadratic (Optimal) Regulator (LQR). The LQR problem is discussed briefly in the next chapter. Section 3.2 shows that reachability always implies controllabiHty, but controllability implies reachability only when the state transition matrix $ of the system is nonsingular. This is always true for continuous time systems, but it is true for discrete-time systems only when the matrix A of the system [or A(k) for certain values of k] is nonsingular. If the system is state reachable, then there always exists an input that transfers any state XQ to any other state xi in finite time.
FIGURE 3.2 A controllable state XQ
In the time-invariant case, a system is said to be reachable (or controllablefrom-the-origin) if and only if its controllability matrix ^ , % = [B,AB,...,A"-l5]G/^"><'^^
(1.1)
has full row rank n, that is, rank % = n. The matrices A G R^^^ and B E R^^"^ determine either the continuous-time state equations X = Ax-\- Bu
(1.2)
or the discrete-time state equations x(k + 1) = Ax(k) + Bu(kl
(1.3)
fc > /^o "= 0. Alternatively, we say that the pair (A, B) is reachable. The matrix % should perhaps more appropriately be called the "reachability matrix" or the "controllability-from-the-origin matrix." The term "controllability matrix," however, has been in use for some time and is expected to stay in use. Therefore, we shall call % the "controllability matrix," having in mind the "controllability-fromthe-origin matrix." We shall now discuss reachability and controllability for discrete-time timeinvariant systems (1.3). If the state x{k) in (1.3) is expressed in terms of the initial vector x(0), then (see Section 2.7) k-\
x(k) = A^x(0) + ^A^-^'^^^Bu{i)
(1.4)
for ^ > 0. It now follows that it is possible to transfer the state from some value x(0) = XQ to some xi in n steps, that is, x{n) = xi, if there exists an n-step input sequence {w(0), w(l),..., u{n - 1)} which satisfies the equation Xi where %n = [B, AB,
A'^XQ
= ^nUn,
(1.5)
A^-i^] = ^ [see (1.1)] and Un = [u^(n - 1), u^(n - 2 ) , . . . , w^(0)]^
(1.6)
From the theory of linear algebraic equations, (1.5) has a solution Un if and only if xi - A'^jco E gi(^),
(1.7)
where 9l(^) = range (%). Note that it is not necessary to take more than n steps in the control sequence since if this transfer cannot be accomplished in n steps, it cannot be accomplished at all. This follows from the Cayley-Hamilton Theorem, in view of which it can be shown that 2/l(^^) = 2/l(^^) for k> n. Also note that 9l(^„) includes ^C^k) iox k < n [i.e., 9l(^„) D 9l(^y^), k < nl (See Exercise 3.1.) It is now easy to see that the system (1.3) or the pair {A,B) is reachable (controllable-from-the-origin), implying that any state xi can be reached from the zero state (XQ = 0) in finite time if and only if rank ^ = n, since in this case 9l(^) = R^, the entire state space. Note that x\ G 9i{%) is the condition for a particular state x\ to be reachable from the zero state. Since 2^(^) contains all such states, it is called the reachable subspace of the system. It is also clear from (1.5) that if the system is reachable, any state XQ can be transferred to any other state xi in n steps. In addition, the input that accomplishes this transfer is any solution Un of
217 CHAPTER 3 :
Controllability, Observability, and Special Forms
218 Linear Systems
(1.5). Finally, depending on xi and XQ, this transfer may be accomplished in fewer than n steps (see Section 3.2). EXAMPLE 1.1. Consider x(k + 1) = Ax(k) + Bu(k), where A
B
Here the controllability (-from-the-origin) matrix ^ is ^ = [B,AB] =
with rank
^ = 2. Therefore the system [or the pair (A, B)] is reachable, meaning that any state xi can be reached from the zero state in a finite number of steps by applying at most n inputs {u(0\ w(l),..., u(n - 1)} (presently, n = 2). To see this, let xi implies that
0 [1
1 u(l) iJWO).
u(l) [u(0)\
-
Then (1.5)
b—a Thus, the control a
M(0) = a,u{V) = b - a will transfer the state from the origin at A: = 0 to the state a b at ^ = 2. To verify this, we observe that x(l) = Ax(0) + Bu(0) = x(2) = Ax(l) + Bu(l)
and
a =
(b-a)
+
Reachability of the system also implies that a state xi can be reached from any other state xo in at most n = 2 steps. To illustrate this, let x(0) = a—2 b~3 1 1 , which will drive the state from at /: = 0 to
plies that xi - A^XQ = L "~ L b-aa-2
11
. Then (1.5) im-
u(l) u(l) . Solving, w(0) u(0) at ^ = 2.
Notice that in general the solution Un of (1.5) is not unique, i.e., there are many inputs which can accomplish the transfer from x(0) = XQ to x(n) = x\, each corresponding to a particular state trajectory. In control problems, particular inputs are frequently selected that, in addition to transferring the state, satisfy additional criteria, such as, e.g., minimization of an appropriate performance index (optimal control). This corresponds to selecting a particular trajectory that, e.g., may result in minimum dissipation of control energy. It is important to remember that reachability and controllability guarantee only the ability of a system to transfer an initial state to a final state by some control input action over a finite time interval. By themselves, reachability and controllability do not imply the capability of a system to follow some particular trajectory. A system [or the pair (A, 5 ) ] is controllable, or controllable-to-the-origin, when any state XQ can be driven to the zero state in a finite number of steps. From (1.5) we see that a system is controllable when A^XQ G 9 l ( ^ ) for any XQ. If rank A = n, a system is controllable when rank ^ = n, i.e., when the reachability condition is satisfied. In this case the nX mn matrix A-""^
=
[A-''B,...,A-^B]
(1.8)
isofinterestandthesystemiscontrollableifandonlyifranA:(A~^^) = rank% = n. If, however, rank A < n, then controllability does not imply reachability (see Section 3.2).
EXAMPLE 1.2. The system in Example 1.1 is controllable (-to-the-origin). To see this, 0 1 u(l) 1 1 a we let xi = 0 in (1.5) and write -A^XQ = = [B,AB] 1 2 b u(0) 1 1 a 1 l]\a 1 r , where Xo = From this we obtain u(0) u(0)\ P. 1 0 1 2 -b 0 -11 la , the input that will drive the state from a at ^ = 0 to -I -i\ L^. —a — b "0^ at ^ = 2. Oj
EXAMPLE 1.3. The system x(k + 1) = 0 is controllable since any state, say, x(0) = , can be transferred to the zero state in one step. In this system, however, the input u does not affect the state at all! This example shows that reachability is a more useful concept than controllability for discrete-time systems. • It should be pointed out that nothing has been said up to now about maintaining the desired system state after reaching it [refer to (1.5)]. Zeroing the input for k^ n, i.e., letting M(fe) = Oforfe> n, will not typically work, unless Axi = xi. In general a state starting at xi will remain at xi for all fc > n if and only if there exists an input u(k), k^ n, such that xi = Axi + Bu(k\ that is, if and only if (/ - A)xi condition may not be satisfied.
^(B).
(1.9)
Clearly, there are states for which this
2. Observability and constructibility In Section 3.3, definitions for state observability and constructibility are given, and appropriate tests for these concepts are derived. It is shown that observability always implies constructibility, while constructibility implies observability only when the state transition matrix O of the system is nonsingular. Whereas this is always true for continuous-time systems, it is true for discrete-time systems only when the matrix A of the system [or when A(k) for particular values of k] is nonsingular. If a system is state observable, then its present state can be determined from knowledge of the present and future outputs and inputs. Constructibility refers to the ability to determine the present state from present and past outputs and inputs, and as such, it is of greater interest in applications. In the time-invariant case a system [or a pair (A, C)] is observable if and only if its observability matrix 0, where C CA
^f^pn>
(1.10)
CA n-l has full column rank, i.e., rank € = n. The matrices A G R^^^ and C E RP^ given by the system description X — Ax + Bu,
= Cx + Du
are
(1.11)
219 CHAPTER 3 :
Controllability, Observability, and Special Forms
220 Linear Systems
in the continuous-time case, and by the system description ^(^ + 1) ^ ^^(y^) _p ^^^^^^
^(^) ^ Cx(yt) + Du(k\
(1.12)
with fe > /:o = 0. in the discrete-time case. We shall now briefly discuss observability and constructibility for the discretetime time-invariant case. As in the case of reachability and controllability, this discussion will provide insight into the underlying concepts and clarify what these imply for a system. If the output in (1.12) is expressed in terms of the initial vector x(0), then k-i
y(k) = CA^x(0) + ^
CA^-^'^^^Bu(i) + Du{k)
(1.13)
i=0
for A: > 0 (see Section 2.7). This implies that y(k) = CA^xo
(1.14)
for /: > 0, where k-l
y{k) = y(k) - ^CA^'^'^^^Bu{i)
+ Du{k)
i=0
for ^ > 0, 3;(0) = yiO) - Du{G), and XQ = x(0). In (1.14), XQ is to be determined assuming that the system parameters are given and the inputs and outputs are measured. Note that if u(k) = 0 for /: ^ 0, then the problem is simplified since y(k) = y(k) and the output is generated only by the initial condition XQ. It is clear that the ability to determine XQ from output and input measurements depends only on the matrices A and C since the left-hand side of (1.14) is a known quantity. Now if ^(0) = XQ is known, then all x(k), ^ > 0, can be determined by means of (1.4). To determine XQ, we apply (1.14) for ^ = 0,.. .,n - 1. Then Yo,n-l
where ©^ = [C^, (CA)^,...,
= ^nXo,
(1.15)
(CA^-^)^]^ - 0 [as in (1.10)] and
hn-i =
[fm...,f(n-l)f.
Now (1.15) always has a solution XQ, by construction. A system is observable if the solution XQ is unique, i.e., if it is the only initial condition that, together with the given input sequence, can generate the observed output sequence. From the theory of linear systems of equations, (1.15) has a unique solution XQ if and only if the null space of 0 consists of only the zero vector, i.e., null (0) = }((€) = {0}, or equivalently, if and only if the only x E R^ that satisfies €x = 0
(1.16)
is the zero vector. This is true if and only if rank € = n. Thus, a system is observable if and only if rank€ = n. Any nonzero state vector x G R^ that satisfies (1.16) is said to be an unobservable state, and J{{€) is said to be the unobservable subspace. Note that any such x satisfies CA^x = 0 for fc = 0, 1 , . . . , n - 1. If rank€ < n, then all vectors XQ that sansfy (1.15) are given by XQ = xop + XQ/Z, where xop is a particular solution and XQU is any vector in }((€). Any of these state vectors, together with the given inputs, could have generated the measured outputs.
To determine XQ from (1.15) it is not necessary to use more than n values for y(k), k = 0,.. .,n - I, ov to observe y(k) for more than n steps in the future. This is true because, in view of the Cayley-Hamilton Theorem, it can be shown that J{(€n) = >r(0^) for k^ n. Note also that J{(€n) is included in Ji(€k) (JV'(O^) C J<(€k)) for k < n. Therefore, in general, one has to observe the output for n steps (see Exercise 3.1). EXAMPLE 1.4. Consider the system x(k + 1) = Ax(k), y(k) = Cx(k), where A = 0 1 C 0 1 and C = [0, 1]. Presently, with rank 0 = 2. Therefore, 1 1 CA 1 1 the system [or the pair (A, C)] is observable. This means that x(0) can uniquely be determined from n = 2 output measurements (in the present cases, the input is zero). 0 1 ^i(O) -1 1 y(0) ^i(O) In fact, in view of (1.15), yiP) 1 1 ^2(0). or ^2(0)J 1 0 Jd). J(l) y{\) - KO)" EXAMPLE 1.5. Consider the system x{k + 1) = Ax{k), y{k) = Cx(k), where A = 1 0 C 1 0 and C = [1,0]. Presently, 0 = with rank 0 = 1 . Therefore, CA 1 1 1 0 the system is not observable. Note that a basis for }((€) is (1.16) implies that all state vectors of the form
which in view of
R, are unobservable. Rela-
1 0 xi(0) . For a solution x(0) to exist, as it must. 1 0 X2(0)J LKDJ we have that y(0) = y(l) = a. Thus, this system will generate an identical output for k> 0. Accordingly, all x(0) that satisfy (1.15) and can generate this output are given by "0" xi(0)l a a + where c G R. _ = c = c 0. .^2(0)J tion (1.15) implies that
y(0)
In general, a system (1.12) [or a pair (A, C)] is constructible if the only vector X that satisfies x = A^x with Cx = 0 for every k > 0 is the zero vector. When A is nonsingular, this condition can be stated more simply, namely, that the system is constructible if the only vector x that satisfies CA~^x = 0 for every fe > 0 is the zero vector. Compare this with the condition CA^x = 0, k > 0, for x to be an unobservable state; or with the condition that a system is observable if the only vector X that satisfies CA^x = 0 for every /: > 0 is the zero vector. In view of (1.14), the above condition for a system to be constructible is the condition for the existence of a unique solution XQ when past outputs and inputs are used. This, of course, makes sense since constructibility refers to determining the present state from knowledge of past outputs and inputs. Therefore, when A is nonsingular the system is constructible if and only if the pnX n matrix CA(1.17) CAhas full rank, since in this case the only x that satisfies CA '^x = 0 for every k ^ 0 is X = 0. Note that if the system is observable, then it is also constructible;
221 CHAPTER3: Controllability, Observability,' and Special Forms
222
Linear Systems
however, if it is constructible, then it is also observable only when A is nonsingular (see Section 3.3). EXAMPLE 1.6. Consider the (unobservable) system in Example 1.5. Since A is non1 0" 1 01 r 1 0^ singular, OA ^ = Since rank OA ^ = 1 < 2, the system [or [-2 Ij .1 OJ .1 0_ the pair (A, C)] is not constructible. This can also be seen from the relation CA ^x = 0,k> 0, that has nonzero solutions x, since C = [1,0] = CA~^ = CA~^ = ••• = CA~^ for /: > 0, which implies that any x =
,c '
7? is a solution.
3. Dual systems Consider the system described by X = Ax + Bu,
y = Cx^Du,
(1.18)
where A G 7?"X^ B G R'''''^, C G RP''^ and D G 7^^^^. The dual system of (1.18) is defined as the system XD = ADXD + BDUD,
yo = CDXD + DDUD.
(1.19)
where AD = A^, BD = C^, CD = B^, and Do = D^. LEMMA 1.1. System (1.18), denoted by {A, B, C, D}, is reachable (controllable) if and only if its dual {A^, BD, CD, DO} in (1.19) is observable (constructible), and vice versa. Proof. System {A, 5, C, D] is reachable if and only if ^ = [5, AB,..., A'^'^B] has full rank n, and its dual is observable if and only if B^ B^A^ \B^{A^Y-^\
has full rank n. Since 0^ = %, {A, B, C, D} is reachable if and only if {AD, ^D, CD, /^D} is observable. Similarly, {A, 5, C, D} is observable if and only if {AD, ^D> CD, DD} is reachable. Now {A, B, C, D] is controllable if and only if its dual is constructible, and vice versa; recall from Sections 3.2 and 3.3, a continuous-time system is controllable if and only if it is reachable; and is constructible if and only if it is observable. For the discrete-time time-invariant case, the dual system is again defined as AD = A^, BD = C^, CD = B^, and DD = D^. That such a system is reachable if and only if its dual is observable can be shown in exactly the same way as in the proof of Lemma 1.1. That such a system is controllable if and only if its dual is constructible when A is nonsingular is true because in this case the system is reachable if and only if it is controllable; and the same holds for observability and constructibility. The proof for the case when A is singular involves the controllable and unconstructible subspaces of a system and its dual. We omit the details. The reader is encouraged to complete this proof after studying Sections 3.2 and 3.3. In the time-varying case, the dual system is defined in a similar manner as given, taking transposes of matrices, and in addition, reversing time. This will not be
discussed further here. We merely wish to point out that the mappings from the original system to the dual system are in this case of the form {A(a + t), B(a + t), C(a + tl D(a + t)} -^ {A^(a - t), C^(a - t), B^(a - t), D^{a - 0}, where a is a fixed real number. Thus, under this transformation we mirror the image of the graph of each function about a point a on the time axis and then take the transpose of each matrix. Note that the mapping between the state transition matrices is of the form ^{a, a^-t)^ ^^{a-t,a). With this definition, it is now possible to establish results similar to Lemma 1.1 for the time-varying case. The proofs involve the Gramians defined in Sections 3.2 and 3.3. Figure 3.3 summarizes the relationships between reachability (observability) and controllability (constructibility) for continuous- and discrete-time systems.
Reachability
Dual
V
v Controllability
Observability
Dual
Constructibility
FIGURE 3.3 In continuous-time systems reachability (observability) always implies and is implied by controllability (constructibility). In discrete-time systems reachability (observability) always implies but in general is not implied by controllability (constructibility). B. Chapter Description This chapter consists of an introduction and two parts. In Part 1, consisting of Sections 3.2 and 3.3, reachability and controllability, and observability and constructibility, are introduced and studied. In Part 2, consisting of Sections 3.4 and 3.5, special forms for state-space representations of time-invariant systems are developed for controllable/uncontrollable, observable/unobservable (continuous- and discretetime) systems. In addition, the poles and zeros of a system are introduced and studied. In the introduction. Subsection 3.1 A, the concepts of reachability (or controllability-from-the-origin) and observability are introduced using discretetime time-invariant systems. The inputs that accomplish the desired transfers of a state are easily derived in terms of the controllability (-from-the-origin) matrix of a system. Conditions for reachability and observability are derived directly in terms of the controllability and observability matrices of the system. Similarly, controllability (or controllability-to-the-origin) and constructibility are also introduced. State reachability and observability are related by duality. Dual systems and the dual notions of reachability (respectively, observability) and controllability (respectively, constructibility) are also discussed. In Section 3.2, reachabihty and controllability are discussed at length for both continuous- and discrete-time systems. In the continuous-time case, the inputs that
223 CHAPTER 3:
Controllability, Observability, and Special Forms
224 Linear Systems
accomplish the desirable state transfers are derived using the reachability (and the controllability) Gramian. Since with only minimal additional work one can treat the time-varying case as well, this is the approach pursued herein, i.e., both time-varying and time-invariant cases are studied. The time-invariant case is discussed separately and can be treated independently of the time-varying case. This adds significant flexibility to the coverage of the material in this chapter. Many criteria for reachability (controllability) are developed. It is shown that reachability implies controllability, and vice versa in the case of continuous-time systems. In discrete-time systems, although reachability implies controllability, controllability does not necessarily imply reachability. This is due to the lack of general time-reversibility in the case of difference equations, as pointed out in Chapter 2. (Note that a detailed section summary is included at the beginning of Section 3.2.) Observability and constructibility are addressed in Section 3.3, in a manner analogous to the treatment of the dual concepts of reachability and controllability in Section 3.2. Both continuous- and discrete-time cases are considered. Observability and constructibility Gramians are used to study these properties in the case of both time-varying and time-invariant continuous-time systems. Once more, the time-invariant case is treated separately and can be studied independently of the more general time-varying case. Observability always implies constructibility in both continuous- and discrete-time systems; however, constructibility always implies observability only in the case of continuous-time systems. This is due to the lack of general time-reversibility of difference equations. (Note that a detailed section summary is included at the beginning of Section 3.3.) In Section 3.4, similarity transformations are used to reduce the state-space representations of time-invariant systems to special forms. First, standard forms for uncontrollable and unobservable systems are developed. These lead to Kalman's Decomposition Theorem and to additional tests for controllability and observability that involve eigenvalues and eigenvectors of the system matrix A (in Subsection B) and to relations between state-space and transfer matrix descriptions (in Subsection C). Controller and observer forms for controllable and observable systems are derived next (in Subsection D). These forms are useful in state feedback control and in state observer design, discussed in Chapter 4. The Structure Theorem is introduced next. This result, which involves the controller (observer) forms and relates the state-space representations to the transfer function matrix of the system, is used in Chapter 5, where state-space reahzations of transfer functions are addressed. In Section 3.5, the poles of a system and of a transfer function matrix are introduced. There, the zeros of the system, the invariant zeros, the input and output decoupling zeros, and the transmission zeros, which are the zeros of the transfer function matrix, are also introduced. The Smith and Smith-McMillan forms of polynomial and rational matrices, respectively, are used to define poles and zeros. Utilizing zeros, one can render certain eigenvalues (system poles) and their corresponding modes unobservable from the output, using state feedback. This leads to the solution of several control problems, such as disturbance decoupling, model matching, and diagonal decoupling. The discussion of poles and zeros of a system {A, B, C, D} and of the corresponding transfer function matrix H(s) also helps to clarify the relationship between internal (state-space) descriptions and external (transfer function matrix) descriptions. This is studied in greater detail in Chapter 5.
C. Guidelines for the Reader
225
Reachability, which is controllability-from-the-origin and controllability (-to-theorigin), together with observability and construetibility are introduced in Subsection 3.1 A using discrete-time time-invariant systems. Careful study of this introductory section leads to early and significant insight into these important system properties, without requiring the mathematical sophistication needed in a careful study of these properties in the continuous-time case. Duality is also discussed in Subsection 3.1 A. In Part 1, reachability and controllability, and observability and constmctibility are introduced in Sections 3.2 and 3.3, respectively, for continuous-time time-varying and time-invariant systems as well as for discrete-time systems. For convenience, detailed summaries of the results with reference to particular definitions and theorems are included at the beginning of these sections. At a first reading, one may concentrate on the time-invariant continuous-time case discussed in Subsections 3.2B and 3.3B. (Recall that an introduction to the time-invariant discrete-time case was presented in Subsection 3.1 A.) The time-invariant case is developed in a self-contained manner in these sections, providing flexibility in coverage of the material. Note that in Corollary 2.12, in Subsection 3.2B, it is shown that the system is reachable if and only if the controllability matrix C has full rank. Theorem 2.13 provides an input u(t) that can accomplish the transfer of the state from a vector value XQ to another vector value xi, provided that such transfer is possible, while Theorem 2.17 gives additional tests for reachability. A relationship between reachability and controllability is established in Theorem 2.16. In an analogous manner, in Corollary 3.8, in Subsection 3.3B, it is shown that a system is observable if and only if the observability matrix € has full rank. A relationship between observability and constmctibility is given in Theorem 3.9, while in Theorem 3.10 additional tests for observability are presented. A useful table of all Gramians used in this chapter is provided in the summary section (Section 3.6). In Part 2, special forms for state-space representations of continuous-time and discrete-time time-invariant systems are introduced. The standard forms for uncontrollable and unobservable representations and the Kalman Decomposition Theorem are presented in Subsection 3.4A, and useful eigenvalue/eigenvector tests for controllability and observability are developed in Subsection 3.4B. The controller and observer forms and the Structure Theorem are discussed in Subsection 3.4D. At a first reading, one could study Subsections 3.4A and 3.4B and cover Subsection 3.4D selectively, concentrating on deriving and using controller and observer forms rather than proofs and properties. Note that the controller and observer forms are used primarily in realization algorithms in Chapter 5, in a method to assign closedloop eigenvalues via state feedback in Chapter 4, and in Chapter 7, to gain insight into the relations between state-space and polynomial matrix representations of linear time-invariant systems. Furthermore, the Structure Theorem discussed in this section introduces polynomial matrix fractional descriptions of the transfer function matrix H(s). These descriptions are very useful in control problems and are discussed further in Chapter 7. In Section 3.5, the poles and zeros of a system are introduced using the Smith form of a polynomial matrix and the Smith-McMillan form of a transfer function matrix H(s). The pole and zero polynomials of H(s) are defined next. This gives rise to the McMillan degree of H(s) and to the order of a minimal
Controllability, Observability, and Special Forms
CHAPTER 3 :
226 Linear Systems
realization, discussed in Subsection 5.2C of Chapter 5. The study of poles and zeros offers significant insight into feedback control systems. At a first reading, Section 3.5 may be omitted without loss of continuity.
PARTI CONTROLLABILITY A N D OBSERVABILITY
3.2 REACHABILITY AND CONTROLLABILITY The objective here is to study the important properties of state controllability and reachability when a system is described by a state-space representation. In Subsection 3.1 A, a brief introduction to these concepts for discrete-time time-invariant systems was given, where it was shown that a system is completely reachable if and only if the controllability (-from-the-origin) matrix ^ in (1.1) has full rank n {rank% = n). Furthermore, it was shown that the input sequence necessary to accomplish the transfer can be determined directly from % by solving a system of linear algebraic equations. In a similar manner, we would like to derive tests for reachability and controllability and determine the necessary system inputs to accomplish the state transfer for the continuous-time case. This is the main objective of this section. We note, however, that whereas the test for reachability in the timeinvariant case {rank ^ = n) can be derived by a number of methods, the appropriate sequence of system inputs to use cannot easily be determined directly from ^ , as was the case for discrete-time systems. For this reason, we use an approach that utilizes ranges of maps, in particular, the range of an important nXn matrix—the reachability Gramian. The inputs that accomplish the desired state transfer can be determined directly from this matrix. However, once this is accomplished, we can develop all the results for the time-varying case as well, with hardly any additional work. This is the approach we will employ. The reader can skip the more general material, however, starting with Definition 2.1, and concentrate on the time-invariant case starting with Definition 2.9, if so desired. The contents of this section are now presented in greater detail. Section description In this section, the concepts of reachability and controllability are introduced and discussed in detail for linear system state-space descriptions. This is accomplished for continuous- and discrete-time systems for both time-varying and time-invariant cases. Reachability for continuous-time systems is discussed first and the reachability Gramian Wr(to, ^i) is defined (in Definition 2.7). It is then shown (in Corollary 2.3) that the system is reachable at t\ if and only if Wr(to, t\) has full rank for some /Q — ^i • Reachability implies that it is possible to transfer the state from a value XQ to some value xi, and system inputs that accomplish this transfer are given in Theorem 2.4 and Corollary 2.5. Controllability is discussed next and the controllability Gramian is defined (in Definition 2.8). It is shown (in Theorem 2.6) that a continuous-time
system is reachable if and only if it is controllable. Two additional results (Theorems 2.8 and 2.9) provide further criteria for reachability and controllability. All these results are then applied to the continuous-time time-invariant case. The above material is presented in a manner that makes possible the study of time-invariant systems, independent of the time-varying case. In Lemma 2.10, a relationship between the reachability Gramian Wr(0, T) and the controllability (-from-the-origin) matrix % = [B, AB,..., A'^'^B] is estabHshed. It is then shown in Corollary 2.12 that a system is reachable if and only if ^ has full rank n. A system input sequence that transfers the state from XQ to xi is derived in Theorem 2.13 and in Corollary 2.14. In Theorem 2.16 a relationship between reachability and controllability is established, and Theorem 2.17 provides additional tests for reachability. For discrete-time systems, in particular, discrete-time time-invariant systems, reachability and controllability are discussed next. Here, the controllability matrix % plays a predominant role. It is shown that when ^ has full rank, the system is reachable (Corollary 2.19), and input sequences that transfer the state to desired values are derived (Theorem 2.20 and Corollary 2.21). In Theorem 2.22 it is shown that reachability in the case of discrete-time systems always implies controllability. In contrast to the continuous-time case, the converse to this statement is generally not true. This is due to a lack of time reversibility in difference equations. When A is nonsingular, then controllability also implies reachability. Finally, in Definitions 2.13 and 2.14 the reachability and controllability Gramians for the discrete-time case are defined for sake of completeness.
A. Continuous-Time Time-Varying Systems We consider the state equation X = A{t)x + B{t)u,
(2.1)
where A{t) G R'''^'', B{t) G R^^^^ and u(t) G R"^ are defined and (piecewise) continuous on some real open interval {a, b). The state at time t is given by x{t, to, XQ) ^ x{t) = ^{t, to)x(to) +
0(^, T)B(T)U(T)
dr,
(2.2)
J to
where ^(t, r) is the state transition matrix of the system, to, t G {a, b), and x(fo) = XQ denotes the initial state at initial time. In the time-invariant case. X = Ax + Bu,
(2.3)
where A G /^^><", B G /^^><'^, (2.2) is still valid with ^{t, T) = ^{t - T, 0) - exp [{t - T)A\ = e^^'-^\
(2.4)
We are interested in using the input to transfer the state from XQ to some other value x\ at some finite time ti > to, [i.e., x(ti) = xi]. Equation (2.2) assumes the form Xi = $(^1, ^o)^0 +
0(^i,
T)5(T)W(T)JT,
(2.5)
227 CHAPTERS:
Controllability, Observability, and Special Forms
228 Linear Systems
and clearly, there exists u(t), t E [/Q, h ] that satisfies (2.5) if and only if such transfer ^f ^^e state is possible. Rewriting (2.5) as ch
jci - (5(^1, ro)xo =
^{ti,
T)B{T)U{T)
dr
(2.6)
and letting xi ^^ x\ - ^{h, to)XQ, we note that the u{t) that transfers the state from XQ at ^0 to x\ at time t\ will also cause the state to reach x\ at ti, starting from the origin at to (i.e., x{to) = 0). For system (2.1), we introduce the following concept. DEFINITION 2.1. A State xi is reachable at time t\ if for some finite to < t\ there exists an input u{t\ t E [to, t\\ that transfers the state x{t) from the origin at ^o, to x\ at time t\ [i.e., that transfers x(t) from x(tQ) = 0 to x(t\) = xi]. Thus, when xi is reachable at ti [with x(to) = 0], then in view of (2.5), there exists an input u such that xi = i' c^(ti,T)B(T)u(T)dr.
(2.7) •
We note that the times ti and to are important individually in the time-varying case only; in the time-invariant case, as is well known by now, ti - to is the important quantity, and typically ^o is taken to be ro = 0 with ti = T,3. finite positive number. The set of all reachable states xi contains the origin and constitutes a linear subspace of the state space (X, R) = (R^, R) (verify this). This gives rise to the following. DEFINITION 2.2. The reachable at ti subspace R[^ of (2.1) is Rr^ = {set of all states x\ reachable at ^i}.
•
When the context is clear and there is no ambiguity, we will write Rr in place ofR',. DEFINITION 2.3. The system (2.1) is (completely state) reachable at ti if every state xi in the state-space is reachable at ti (i.e., Rr = R'^). In this case, we equivalently make reference to reachable pair (A(t), B(t)) atti. • A reachable state is sometimes also called controllable-from-the-origin. Additionally, there are also states defined to be controllable-to-the-origin or simply controllable. In particular, we have the following notion. DEFINITION 2.4. A state XQ is controllable at time to if for some finite t\ > to there exists an input u(t\ t E [to, ti] that transfers the state x(t) from xo at ^o to the origin at time ti [i.e., from x(to) = xo to x(ti) = 0]. • In view of (2.5), there exists an input u such that -^(tu
to)Xo = f ' ^(h,
T)B{T)u{T)dT,
(2.8)
JtQ
or by premultiplying by ^~^{t\, to) = ^(to, ti) (see Section 2.3), - x o = f ' 0(^0, T)B{T)u{r)dT, where the semigroup property ^{to, t\)^{t\y r) = 0(^0, r) was used.
(2.9)
Similar to the case of reachable states, the set of all controllable states includes the origin, and is a linear subspace Re of the state-space X (i.e.. Re C X).
229 CHAPTERS:
Controllability, Observability, • and Special Forms
DEFINITION 2.5. The controllable at to subspace /?J? of (2.1) is R^c = {set of all states XQ controllable at ^o}It is denoted by Re for convenience when there is no ambiguity.
DEFINITION2.6. The system (2.1) is (completely state) controllable at to if every state xo in its state-space is controllable (i.e., if Re = R^). In this case, we equivalently make reference to controllable pair (A(t), B(t)) at ^. • Discussion Relation (2.7) shows that for x = A(t)x + B(t)u given in (2.1) and for given ti, the range of the integral map L = L(u, to, ti) =
^{t\,
T)B{T)U{T)
dr
(2.10)
with u{t), t E [^0, ^i], and with t^ varying over all finite values ^Q < ^i, is exactly the reachability subspace Rr, since a state x\ is reachable if there exists a t^ and u such that xi E 2/l(L). Notice that in view of (2.6), the input u which transfers the state from the origin at t^io x\ at t\ also transfers the state from XQ at t^io xi at t\, where xi = x\ — ^{t\, ^)xo. For fixed XQ, since {x\} spans the reachability subspace /^^\ this relation yields all states xi that can be reached from XQ in finite time t\ - to-
In Lemma 2.1, the range of L is shown to be equal to the range of a matrix, the reachability Gramian Wr(to, t\), which is rather easy to determine. In the timeinvariant case, it is also shown to be equal to the range of the controllability matrix %. Before proving these results, the relation between reachability (controllabilityfrom-the-origin) and controllability (controllability-to-the-origin) is discussed; the exact relation is proved in Theorem 2.6. In view of (2.7) and (2.8), a vector x is reachable (controllable-from-the-origin) at t\ if there exists a finite ^o and u(t), t E [^o, ^i], so that x E 9l(L), where L is defined in (2.10), and it is controllable (-to-the-origin) at ^o if 0(^i, to)x E Sl(L). It is shown later (Theorem 2.6) that the system (2.1), or the pair (A, B), is (completely state) reachable if and only if it is controllable. This is the reason why only one term is typically used in the literature when describing these properties for continuoustime systems. For discrete-time systems, however, the situation is different. In this case, as will be shown later in this section, if the pair (A, B) is reachable, then it is also controllable, but not necessarily vice versa; that is, in the discrete-time case controllability does not necessarily imply reachability. Indeed, controllability implies reachability only when the state transition matrix ^(k, ko) has full rank, which is not always true in discrete-time systems. As discussed in Chapter 2, this is due to the lack of the "time reversibility" property. On the other hand, in the case of continuous-time systems, ^(t, r) is always nonsingular. In such systems, reachability implies that any state xi can be reached from any other state XQ in finite time ti - to. This property is sometimes used in the literature to define "controllability." An input that achieves this transfer is given later in Corollary 2.5. In the discretetime systems literature, the term that is typically used is "reachability"; however, for simplicity, the term "controllability" is sometimes also used, with some sacrifice of
230 Linear Systems
accuracy. We will use both terms, reachability and controllability, with a warning to the reader when use of the term controllability (-from-the-origin) is made, instead of reachability. Now suppose there exists an input u which transfers the state of the system from x{to) = 0 to x{ti) = x i , that is, (2.7) is true. The integral in (2.7) is a map L = L{u,to,ti) defined in (2.10) that maps an input u{t) ^ R^ defined over [^o^^i] to states xi ^ R^. We are interested in the range of L, ^ ( L ) since it contains all the states that can be reached from the origin, x{to) = 0, at time ti, by varying the input u. Note that L has infinite-dimensional domain, and therefore it is not easy to determine its range directly. In the following we show that ^ ( L ) is equal to the range of an important matrix, the reachability Gramian. D E F I N I T I O N 2.7. The reachability Gramian of the system x = A(t)x-\-B(t)u nxn matrix Wr{toA)
= rO(n,T)5(T)5^(T)0^(n,T)JT, Jto
where 0(r, T) denotes the state transition matrix.
is the (2.11)
•
Note that Wr is symmetric and positive semidefinite for every ti > to; that is, Wr = W^ and Wr > 0 (show this). Now let to < ti be given. Then the following lemma can be shown. L E M M A 2.1. ^{L{u,to,ti))
=^{Wr{to,ti)).
Proof, We first show that ^(W^) C ^ ( L ) . Let xi G ^(W^); that is, there exists 7]i eR^ such that WrT]i =xi. Choose UI{T) = B^(T)^^{ti,T)rii. Then L{ui,to,ti) = Jl^ 0{ti,T)B{T)B^
{T)0^ {ti,T)dT\rii
=Wrrii =xi. Therefore, xi G ^ ( L ) , and since
xi is arbitrary, it follows that ^{Wr) C ^ ( L ) . We shall now show that ^ ( L ) C ^(Wr), which together with ^(Wr) C ^ ( L ) proves that ^ ( L ) = ^(W^). Let xi G ^ ( L ) , i.e., there exists an input ui such that L{ui,to,ti) = xi. We assume that xi ^ ^(W^) and we shall show that this leads to a contradiction. This implies that the null space of W^(^0,^1) is nonempty. Wr is symmetric, and so the range of Wr is the orthogonal complement of its null space (prove this). Thus for any u G ^{Wr) and v G yK(Wr), u^v = 0. Also, we may write xi = x^^ +x'/ with x[ G ^(Wr) and x'l G yK{Wr) (x'l ^ 0 since xx ^ ^(Wr)). Then there exists X2 G yK(Wr) such that X2x'( ^ 0, which implies X2X1 7^ 0. Now ^2 Wr(^0,^1 )-^2 = 0 = Ho [ 4 ^ ( ^ i ' ^ ) ^ ( ^ ) ] [xl^{tuT)B{T)fdT = g' II x^O(ri,T)5(T) 11^ J T , which shows that^2 0(^1, T)B{T) = 0 for every T G [^0,^1]• This in turn implies that^2xi = X2L{u\) = Jl^[xl^{tuT)B{T)]ui{T)dT = 0, which is a contradiction since ^2X1 7^ 0. Therefore, XI G ^(Wr), which implies that ^ ( L ) C ^(Wr). • Lemma 2.1 shows that the set of all states that can be reached at time ti from the origin at some finite time to < ti, is given by ^(Wr{to,ti)), the range of the reachability Gramian. THEOREM 2.2. Consider the systemi = A(^)x + 5(^)w givenin (2.1). There exists an input u that transfers the state to xi at ti from the origin at some finite time ^0 < ^1 ^ if and only if there exists finite time ^0 < h so that XI
e^{Wr{to,ti)).
Furthermore, an appropriate u that will accomplish this transfer is given by
231 (2.12)
u(t) = B^{t)^^{ti, t)j]i with 171 a solution of Writ^, ti)ri\ = xi and t G [to, ti].
Proof, In view of Lemma 2.1 and the definition of L(u) in (2.10), the proof of the first part of the theorem is straightforward. To prove the second part of the theorem, note that (2.12) was used in the proof of Lemma 2.2 to accomplish the transfer to xi. • COROLLARY 2.3. The system x = A(t)x + B(t)u is (completely state) reachable at ti, or the pair (A(t), B(t)) is reachable at ti, if and only if there exists finite to < t\ such that (2.13)
rank Wr{to, t\) = n.
Proof. In view of Theorem 2.2, all states xi can be reached at ti if and only if for some to < h,^{Wr(toyt\)) = ^", the entire state space. This is true if and only if rank Wr{to, ti) = n for some finite ^ < ^1. • The following result is useful in accomplishing the transfer from a state XQ to another state x\ in some given finite time t\ -to. THEOREM 2.4. There exists an input u that transfers the state of the system x = A(t)x + B(t)u from xo at time to to xi at time ti > to if and only if Xi - ^(ti, to)Xo E ^(Writo,
(2.14)
h)).
Furthermore, such input is given by u(t) = B^(t)^'^(ti, t)r]i
(2.15)
with 171 a solution of W'^(^o, ^i)'»7i = -^1
(2.16)
-^{ti,to)xo'
Proof The proof is straightforward in view of Theorem 2.2 and the fact that there exists an input which transfers the state from xo at to to xi at t\ if and only if it transfers the state from the origin at to to xi = x\ - 0(fi, to)xo at t\ [see (2.6)]. • - 1 e^' B{t) = 0 0 -1 The state transition matrix was calculated in Example 3.4, Section 2.3 (of Chapter 2), to be
EXAMPLE 2.1. Consider i: = A(Ox + 5(0w, where A(0 =
O-it-T)
^{U T )
Here ^{t,
T)B(T)
0
W^'— e 7+3x1
i^z.^
-a-r)
and the reachability Gramian of the system is
WritoJl) to L
0
dr =
'(ti - to)e-^' 1 0'
0
0_
It is clear that rankWr{toy ti) < 2 = /2 for any to < t\ and therefore the system is not reachable at t\. Note that since t\ is arbitrary, the system is not reachable at any finite time. However, the state can be transfered from the origin to a state \(X
xi E ^{Wr{to, h)). In particular, in view of Theorem 2.2, let xi =
I, a E R,
CHAPTER 3 :
Controllability, Observability, and Special Forms
232 Linear Systems
-to
and solve Wr{to, ti)r]i = xi to obtain r/i =
where p G R arbitrary. Then, _e^h
in view of (2.12), u(t) = [<^(ti,t)B{t)fr]i
= [^"^i,0]
-^0
-e^^ will
to
drive the state from the origin at ^o to xi at ti. To verify this, we note that x(ti) = \;'^{h,T)B{T)u{T)dT
=
0
-e'^ih -to)
= ,,_. Notice that for the transfer
t\ — to
to be accomplished in a short period of time, t\ - to = e with e small, the required control magnitude can be quite large since uit) = (a/e)e^^. • The last observation in Example 2.1 points to two important aspects that we now elaborate on. First, we note that the faster the state of the system is required to move (the smaller the ti - to = e) and the further away the desired state x\ is (the larger the a ) , the larger the required control magnitude will be. This makes intuitive sense since it simply states that the more sudden and drastic the change in the state, the larger the required control force will be (think, e.g., of a simple mechanical spring system). Second, it is clear that the property of reachability (controllability) implies the ability to change the state of the system very fast indeed, paying for this of course in terms of increased control magnitude (see Example 2.1 and Exercise 3.12). Intuitively, this is not always possible in the case of physical processes, where only limited control action is typically available. This points to some of the limitations of linear system models that do not include information about input saturation limits, nonlinear behavior, limitations of output sensors, and the like. EXAMPLE 2.2. Consider the system described by i = A(t)x + B(t)u, where A(t) = -1 e^' . The state transition matrix ^(?, T) is given in Example 2.1. B(t) = 0 -1 -t+2T-\
and the reachability Gramian Wr(to, t\)is such that e~ the system is reachable at ti (show this). •
Here 0(?, r)B(T) =
The following result demonstrates the importance of reachability in determining an input u to transfer the state from any XQ to any x\ in finite time. COROLLARY 2.5. Let the system i: = A(t)x + B(t)u be (completely state) reachable at time t{, or let the pair (A(t), B(t)) be reachable at ti. Then there exists an input that will transfer any state xo at some finite time to < ti, to any state xi at time ti. Such input is given by u(t) = B^(t)^^(ti,t)W;\to,ti)[xi
-<^(ti,to)xo]
(2.17)
fort E [^0,^1].
Proof, In view of Corollary 2.3, reachability implies that, given ti, rank Wr(to, ti) = n for some ^o < t\ or that ^{Wr{to, t\)) = /?", the whole space, for some to. This implies that any vector xi - ^ ( ^ i , to)xo G ^(W;.(^, ^0), which in view of Theorem 2.2 and (2.12) implies that the input in (2.17) is an input which will accomplish this transfer. • There are many different control inputs u that can accomplish the state transfer from Xo at ^ = to to x\ at t = ti. It can be shown that the input u given by
(2.17) accomplishes this transfer while expending a minimum amount of energy. In particular, among all the control inputs u(t) that will transfer the state from XQ at to to xi at t\, u(t) in (2.17) minimizes the cost functional J/^ ||W(T)|P J r , where \\u(t)\\ = [w^(Ow(0]^^^theEuchdean norm of u(t). We shall now establish a connection between controllability and reachabihty of the continuous-time system x = A(t)x + B(t)u. THEOREM 2.6. If the system x = A(t)x + B(t)u, or the pair (A(t), B(t)), is reachable at ti, then it is controllable at some ^ < ^i. Also, if it is controllable at ^o, then it is reachable at some ti > to. Proof. It was shown in Corollary 2.3 that for reachability of (2.1) at ti, we must have rank Wr(to, ti) = n. A similar test for controllability can be derived in an identical manner. In particular, in view of (2.9), it is clear that the range of L = L{u,tQ,ti) =
(^(to, T)B(T)U(T)
(2.18)
dT
is of (present) interest [compare with L in (2.10) used to prove reachability]. A result similar to Lemma 2.1 can now be established using an identical approach, namely, that ^(L(u, to, ti)) = '3i(Wc(to, ^i)), where Wdto, ti) is the controllability Gramian, defined next. • DEFINITION2.8. The controllability Gramian of the system x nX n matrix Wc(to,ti)
^(to,
T)B{T)B^{T)^^{to,
A(t)x + B(t)u is the
(2.19)
T)dT,
where ^{t, r) denotes the state transition matrix. Continuing the proof of Theorem 2.6, we note that it can be shown that the input uiit) =
(2.20)
-B^(t)^^(toJ)vi
with 171 such that Wdto, ^1)171 = ^0 satisfies L(wi, to, ti) = -XQ or relation (2.9). Thus, 171(0 drives the state from xo at time to to the origin at time ti > to (compare with Theorem 2.2). As in Corollary 2.3 for the case of reachability, it can be shown in an analogous manner that the system is (completely state) controllable at to if and only if there exists ti > to so that (2.21)
rank Wdto, ti) = n. Next, we note that in view of the definitions of Wr and Wc, Wr(to,ti)
=
(2.22)
^(tiJoWc(to,ti)^^(tuto)-
Since 0(ri, ^o) is nonsingular for every to and t\, rank Wr(to, ti) = rank Wdto, ti) for every to and ti. Therefore, the system is reachable if and only if it is controllable. • EXAMPLE2.3. Consider the system x lability Gramian is given by Wc{to,ti) to
0
A{t)x + B{t)u of Example 2.1. The control-
[e-'^0]dT
=
(ti-to)e-^'o 0
0 0
Compare this with the reachability Gramian of Example 2.1 and note that 4>(ri, to)) (ti - to)e-^'^+'o) 0 to) 'dti Wc(to,ti)<^^(ti,to) ^^(^1,^0) = 0 0 0 Wr(to, ti), as expected [see (2.22)].
233 CHAPTER 3 :
Controllability, Observability, and Special Forms
234 Linear Systems
Before proceeding, we note that a relation similar to (2.17) can be derived using ^^^ Gramian Wdto, ti) and (2.20) in place of Wr(to, ti). In particular, an appropriate input that transfers the state from XQ at to to xi at ti is given by u(t) = -B^(t)^^(to,
t)W;\to,
ti)[xo - $ ( % h)xil
(2.23)
We ask the reader to show that this relation can also be derived from (2.17), using (2.22). Additional criteria for reachability and controllability First recall from Chapter 2 the definition of a set of linearly independent functions of time and consider in particular n complex-valued functions fi(t\ i = 1 , . . . , /i, where f^{t) G C^. Recall that the set of functions fi,i = 1 , . . . , n, is linearly dependent on a time interval [ti, ^2] over the field of complex numbers C if there exist complex numbers at, i = \,.. .,n, not all zero, such that ^ i / i ( 0 + • • * + ^nfn(t) = 0
for all t in [ti, ^2];
otherwise, the set of functions is said to be linearly independent on [ti, ^2] over the field of complex numbers. It is possible to test linear independence using the Gram matrix of the functions fi. LEMMA 2.7. Let F(t) G C^"^ be a matrix with fi(t) G C^^"" in its /th row. Define the Gram matrix of fi(t), i = 1,..., ^, by W(tiJ2) =
F{t)F\t)dt,
(2.24)
where ( • )* denotes the complex conjugate transpose. The set fi{t\ i = I,.. .,n,is linearly independent on [ti, ^2] over the field of complex numbers if and only if the Gram matrix W(ti, ^2) is nonsingular, or equivalently, if and only if the Gram determinant det W{ti, t2) 7^ 0. Proof, {Necessity) Assume the set fi,i = I,..., n, is linearly independent but W{t\, ^2) is singular. Then there exists some nonzero a ^ C^^"^ so that a W^(ri, ti) = 0, from which aW(ti,t2)a* = l^[\aF(t))(aF(t)ydt = 0. Since (aF(t))(aF(t)y > 0 for all r, this implies that aF(t) = 0 for all t in [^1, ^2], which is a contradiction. Therefore W(ti, ^2) is nonsingular. (Sufficiency) Assume that W(ti, ^2) is nonsingular but the set fi,i = 1,..., n, is linearly dependent. Then there exists some nonzero a ^ C^^^ so that aF(t) = 0. Then aW(ti, ^2) = J/^ aF(t)F*(t)dt = 0, which is a contradiction. Therefore the set fi, i = 1,..., /t, is linearly independent. • We will use the above result to derive additional tests for reachability and controllability in this section and for observability and constructibility in the next section. In the following two theorems, we repeat some earlier results, for convenience. THEOREM 2.8. The system x = A(t)x + B(t)u is (completely state) reachable at ti (i) if and only if there exists finite ^0 < ^1, such that rankWritoJi) = n,
(2.25)
),^l) ^- j / ^\,l'^(tuT)l where Wr(to,ti) (ri, T)5(T)B^(T)^^(ri, T)(iT, the reachability Gramian, or equivalently,
(ii) if and only if there exists finite ^o < ^i, such that the n rows of ^{h, t)B{t)
235
(2.26)
are Hnearly independent on [t^, t\\ over the field of complex numbers. Proof. Part (i) was established in Corollary 2.3, while part (ii) is a direct consequence of the previous lemma and the definition of the reachability Gramian. • Similar results can be derived for controllability. Specifically, we have the following result. THEOREM 2.9. The system x = A(t)x + B(t)u is (completely state) controllable at to (i) if and only if there exist finite ti > to such that rank Wdto, ti) = n,
(2.27)
where Wc(^, ti) = j^^^ 0(^0, T ) 5 ( T ) 5 ^ ( T ) 0 ^ ( ^ , T) (ir, the controllability Gramian, or equivalently, (ii) if and only if there exists finite ti > to such that the n rows of 0(^, t)B(t)
(2.28)
are linearly independent on [^o, ^i] over the field of complex numbers. Proof, The proof is analogous to the proof for Theorem 2.8.
•
Notice that premultiplication of $ ( ^ , t)B(t) by the nonsingular matrix 0(?i, ^o) yields $(fi, t)B(t) [refer to (2.26) in the reachability theorem; compare with (2.22)]. This can be used to prove in an alternative way the result of Theorem 2.6, that reachability (controllability-from-the-origin) implies and is implied by controllability (-to-the-origin), in the case of continuous-time systems (show this).
B. Continuous-Time Time-Invariant Systems We shall now apply the results developed above to time-invariant systems x = Ax + Bu given in (2.3). In this case, the state transition matrix
CHAPTER 3 :
Controllability, Observability, and Special Forms
236 Linear Systems
(ii) (iii)
The set of all controllable states Re is the controllable sub space of the system i = Ax + Bu, or of the pair (A, 5). The system i = Ax+Bu, or the pair (A, 5), is {completely state) controllable if every state is controllable, i.e., if Re = R^. •
DEFINITION 2.11. The n X n reachability Gramian of the time-invariant system x= Ax-\-Bu is Wr{0, T) ^ r e^^-'^^BB^e^^-'^'^'dT. Jo
(2.29) g
Note that Wr is symmetric and positive semidefinite for every T > 0, i.e., W^ = Wj and Wr > 0 (shovv^ this). Let the n x mn controllability (-from-the-origin) matrix (or more precisely, the reachability matrix) be ^=[5,A5,...,A^-^5],
(2.30)
and recall that ^ vv^as also defined in Section 3.L It is now shovv^n that in the time-invariant case the range of Wr(0, T), denoted by ^ ( W r ( 0 , r ) ) , is independent of T, i.e., it is the same for any finite T{> 0), and in particular, it is equal to the range of the controllability matrix ^. Thus, the reachable subspace Rr of a system is given by the range of ^,^(^), or the range of Wr{0, T),^{Wr{0, r ) ) , for some finite (and therefore for any) T>0. LEMMA 2.10. ^ ( W , ( 0 , r ) ) = ^ ( ^ ) for every T > 0. Proof, We first show that ^(Wr) C ^ ( ^ ) for some T > 0. Let xi e ^{Wr) for some r > 0. In view of Lemma 2.1, xi G ^ ( L ) , i.e., there exists ui such that L(wi,0,r) = /Q {Qxp[(T — T)A]}Bui{T)dT = xi. Using the series definition exp[A^] = Er=o(^ /^-M ' -^1 ^^^ t)e written as xi =
Yk=o'^^
f^{iT-Tf/kl)ui{T)dT\
or, in
view of the Cay ley-Hamilton Theorem, xi = 2^~Q A^5ay^(r), where ak{T) is appropriately defined. This imphes thatxi G ^ ( ^ ) . Since xi is arbitrary, ^{Wr) C ^ ( ^ ) . We shall now show that ^ ( ^ ) C^{Wr).LQtxi G ^ ( ^ ) , i.e., there exists 7] i eR"""^ such that ^ r | i = xi. Assume that xi ^ ^(W^) for some 7 > 0. We shall show that this leads to a contradiction. This implies that the null space of Wr is nonempty. Wr is symmetric, and so the range of Wr is the orthogonal complement of its null space (prove this). Thus for any u e ^(Wr) and v G yK{Wr),u^v = 0. Also, we may write XI = x[ +x'/ with x[ G ^{Wr) and x'l G ^ ( W , ) (x'l ^ 0 since xi ^ ^(Wr)). Then there exists X2 G yl^{Wr) such that X2x'( ^ 0, which implies X2X1 7^ 0. Next, consider xlWr{0,T)x2 = 0 = /o^[x^{exp[(r - T)A]}5][x^{exp[(r - T)A]}B]^dT = JQ \\xl{Qxp[{T - T)A]}B\\ldT, which shows that x^exp[(r - T)A]B = 0 for every T G [0, r ] . Taking derivatives of both sides with respect to T and evaluating at T = T, we obtain x^B = -x^AB = --- = {-ifx^A^B = 0 for every ^ > 0. Thus, x^A^B = 0 for every ^ > 0, and therefore, ^2X1 = X2^r\\ = 0, which is a contradiction since x^xi y^ 0. Therefore, xi G ^(W^), which implies that ^ ( ^ ) C ^(W^). This, together with ^(Wr) C ^ ( ^ ) , shows that ^{Wr) = ^ ( ^ ) . • Lemma 2.10 shovv^s that given the time-invariant system x = Ax-\-Bu, if x(0) = 0, then the set of all states that can be reached in finite time, i.e., the reachability subspace Rr is given by ^(^), the range of the controllability matrix, or equivalently, by ^ ( W r ( 0 , r ) ) , the range of the reachability Gramian, vv^here T > 0 is any finite time.
EXAMPLE 2.4. For the system i: = Ax + Bu v^iih A = 1 0
t and e^^B = 1
[T - T,l]dT
=
(T -rf T -T
0 0
1 andB 0
. The reachability Gramian is ^^^(0, T) = T - 7 dr = 1
, we have
CHAPTER 3 :
-T
T
1
. Since det W,(0,7) =
^ r ^ 7^ 0 for any T > 0, ra^/: W,(0, T) = ^ and (A, 5) is reachable. Note that % 0 1 and that ^(WM T)) = S?l(^) = R^, as expected (Lemma 2.10). [5, AB] = 1 0 1 1 0 instead of and (A, B) is not reachthen ^ = [B, AB] If B -- 0 0 0 T 1 0 able. In this case e^^B = and the reachability matrix is Wr(0, T) = \^ dr = 0 0 T 0 Notice again that = ^(Wr(0, T)) for every T > 0. 0 0 12
THEOREM 2.11. Consider the system x = Ax -\- Bu and let x(0) = 0. There exists input u that transfers the state to xi in finite time if and only if xi G S/l(^), or equivalently, if and only if xi G ^(Wr(0, T)) for some finite (and therefore for any) T. Thus, the reachable subspace Rr = S/l(^) = ^(Wr{0, T)). Furthermore, an appropriate u that will accomplish this transfer in time T is given by
u(t) = ^ V ^ ^ - ^ i
(2.31)
with r/i such that Wr(0, T)r]i = xi and t G [0, T]. Proof. Apply Theorem 2.2 to the time-invariant case and then use Lemma 2.10.
•
Note that in (2.31) no restrictions are imposed on time T, other than that T be finite. T can be as small as we wish, i.e., the transfer can be accomplished in a very short time indeed. COROLLARY 2.12. The system x = Ax + Bu, or the pair (A, 5), is (completely state) reachable, if and only if rank"^ = n,
(2.32)
rankWr(0,T) = n
(2.33)
or equivalently, if and only if
for some finite (and therefore for any) T. Proof, Apply Corollary 2.3 to the time-invariant case and use Lemma 2.10.
237
•
THEOREM2.13. There exists input u that transfers the state of the system x = Ax+Bu from XQio x\ in somefinitetime T if and only if XX - e^^XQ G m.{%),
(2.34)
xi - e^^XQ G ^{WXO, T)).
(2.35)
or equivalently, if and only if
Controllability, Observability, and Special Forms
238
Such input is given by
Linear Systems
u(t) = B^e^^^^-'^i]i
(2.36)
with t E [0, r ] , where r/i is a solution of Wrifi, T)7]i = xi-
e^^XQ.
(2.37)
Proof, Apply Theorem 2.4 to the time-invariant case and use Lemma 2.10.
•
The above leads to the next result, which establishes the importance of reachability in determining an input u to transfer the state from any XQ to any xi in finite time. COROLLARY 2.14. Let the system i: ^ Ax + 5M be (completely state) reachable, or the pair (A, B) be reachable. Then there exists an input that will transfer any state XQ to any other state x\ in some finite time T. Such input is given by u{t) = B^e^^^^-'^W;\0,T)[xi
(2.38)
- e^^x^-]
for r e [0, r ] . Proof, This result is the time-invariant version of Corollary 2.5. In view of Corollary 2.12, reachability implies that ra^y^W;-(0,r) = ^ for some T or that 2?l(W^(0, T)) = /?", the whole state space. This implies that any vector xi - e^-^xo E '3i(Wr(0, T)) that, in view of Theorem 2.13, implies that the input in (2.38) is an input which will accomplish this transfer. • There are many different control inputs u that can accomplish the transfer from xo to xi in time T. It can be shown that the input u given by (2.38) accomplishes this transfer while expending a minimum amount of energy; in fact, u minimizes the cost functional j ^ ||W(T)|P J r , where \\u(t)\\ = [u^(t)u(t)y^^ denotes the Euclidean norm of u(t). 0 1 and 5 is reachable 0 0 (see Example 2.4). A control input u(t) that will transfer any state XQ to any other state Xi in some finite time T is given by (see Corollary 2.14 and Example 2.4)
EXAMPLE 2.5. The system x = Ax + Bu with A
u(t) = B^e'^'^^^-'^W;\0,
T)[xi
XQ]
12
= [T-t, 1]
J2
6
4
j2
J
Xi
Xo
EXAMPLE 2.6. For the (scalar) system x = -ax + bu, determine u(t) that will transfer the state from x(0) = XQ to the origin in T sec; i.e., x{T) = 0. We shall apply Corollary 2.14. The reachability Gramian is Wr(0, T) =
-[1 2a
[e'
~^''^]. Note [see (2.41) below] that the controllability Gramian is Wc(0, T) 1]. Now in view of (2.38), we have
^(0 = be-^^-'^''
2a
FT
-2aT
239
[-e-^'^x^]
CHAPTER 3 :
2a e-^'^ e'^XQ -laT b 1 2a 1 ^"'Xo. b ^2«r _ I
Controllability, Observability, and Special Forms
To verify that this u{t) accompHshes the desired transfer, we compute x{t) = ^^^xo + ''e^^'-^^Bu(T)dT = e- XQ + Jo e-'''e''^bu{r)dT = e'^\x^ + JQ e'^'b X 2a
1
^2at ,
dr = e
1 -
1
XQ. Note that x{T) = 0, as desired, and
also that x(0) = XQ. The above expression shovv^s also that for t > T, the state does not remain at the origin. An important point to notice here is that as T ^ 0, the control magnitude \u\^ oo. Thus, although it is (theoretically) possible to accomplish the desired transfer instantaneously, this will require infinite control magnitude. In general, the faster the transfer, the larger the control magnitude required. • We shall novs^ establish the relationship between reachability and controllability for the continuous-time time-invariant systems (2.3). Applying (2.8) to the time-invariant case, XQ is controllable v\^hen there exists u(t), t G [0, Tl so that -e^^XQ
or when e^^xo £ ^(Wr(0,
=
T)), or equivalently, in view of Lemma 2.1, when e^^XQ G 2/l(^)
(2.39)
for some finite T. Recall that xi is reachable when XX G 3 l ( ^ ) .
(2.40)
We require the following technical result. LEMMA 2.15. If X E ^{%), then Ax G S/l(^); i.e., the reachable subspace Rr = ^{%) is an A-invariant subspace. Proof, If X G S^(^), this means that there exists a vector a such that [B, AB,..., A«-i5]a - X. Then Ax = [AB, A^B,..., A"5]a. In view of the Cay ley-Hamilton Theorem, A" can be expressed as a linear combination of A"~^ . . . , A, / , which implies that Ax = %fi for some appropriate vector jS. Therefore, Ax G S^(^). • THEOREM 2.16. Consider the system x = Ax + Bu. (i) A state x is reachable if and only if it is controllable, (ii) Re = Rr. (iii) The system (2.3), or the pair (A, B), is (completely state) reachable if and only if it is (completely state) controllable. Proof, (i)Letxbe reachable; that is, xG2?l(^).Premultiplyxby^^^ = X1=o(T^/kl)A^ and notice that in view of Lemma 2.15, Ax, A^x,..., A^x G 2?l(^). Therefore, e^^ x G 2/l(^) for any T, which, in view of (2.39), implies that x is also controllable. If now x is controllable, i.e., e^^x G ^ ( ^ ) , then premultiplying by e~^^, the vector e~^^ [e^^x) = X will also be in ^(%). Therefore, x is also reachable. Note that the second part of (i), that controllabihty implies reachability, is true because the inverse {e^'^Y^ = e~^^ does
240 Linear Systems
exist. This is in contrast to the discrete-time case where the state transition matrix 4>(/c, 0) ^^ nonsingular if and only if A is nonsingular [nonreversibihty of time in discrete-time systems (see Section 2.7)]. Parts (ii) and (iii) of the theorem follow directly from (i). • The reachability Gramian for the time-invariant case, Wr(0, T), was defined in (2.29). Similarly, in view of Definition 2.8, we make the following definition. DEFINITION2.12. The controllability Gramian in the time-invariant case is the nXn matrix rT
Wc(0,T)^\
Jo
e-^'BB^e-^^'dr.
(2.41) •
We note that which can be verified directly [see also (2.22)]. As was done in the time-varying case above, we now introduce a number of additional tests for reachability and controllability of time-invariant systems. Some earlier results are also repeated here for convenience. THEOREM 2.17. The system x = Ax + Bu'\^ reachable (controllable-from-the-origin) (i) if and only if rank Wr(fi, T) = n
for some finite T > 0,
rT
where
WM T) =
e^^-''^^BB^ e^'^'^^^^ dr,
(2.42)
the reachability Gramian, or (ii) if and only if the n rows of (2.43)
are linearly independent on [0, ^) over the field of complex numbers, or alternatively, if and only if the n rows of {sI-Ay^B
(2.44)
are linearly independent over the field of complex numbers, or (iii) if and only if rank% = n,
(2.45)
where % = [B,A,B,..., A"~^B], the controllability matrix, or (iv) if and only if rank [sil - A,B] = n
(2.46)
for all complex numbers st, or alternatively, for si, i = \,.. .,n, the eigenvalues of A. Proof, Parts (i) and (iii) were proved in Corollary 2.9. In part (ii), rank W^(0, T) = n implies and is implied by the linear independence of the n rows of e^^'^^^B on [0, T] over thefieldof complex numbers, in view of Lemma 2.7, or by the linear independence of the n rows of e^^B, where t = T -t, on [0, T]. Therefore the system is reachable if and only if the n rows of e'^^B are linearly independent on [0, oo) over the field of complex numbers. Note that the time interval can be taken to be [0, ^) since in [0, 7], T can be taken to be anyfinitepositive real number. To prove the second
part of (ii), recall that ^{e^^B) = (si — A)~^B and that the Laplace transform is a one-to-one linear operator. Part (iv) will be proved later in this chapter, in Corollary 4.6. • Results for controllability that are in the spirit of those given in Theorem 2.17 can also be established. The reader is asked to do so. This is of course not surprising since it vv^as shovv^n (in Theorem 2.16) that reachability implies and is implied by controllability. Therefore, the criteria developed in the theorem for reachability are typically used to test the controllability of a system. 0 1 and 5 : 0 0 Example 2.4), we shall verify Theorem 2.17. The system is reachable since
EXAMPLE 2.7. For the system x = Ax-\- Bu, where A
\T^
(as in
\T^-
(i) the reachability Gramian Wr(0,7):
has rankWr{0,T) •• 2^
for any 7 > 0, or since
has rows that are linearly independent on [0,oo) over the field
(ii) e^^B •
of complex numbers (since ai -1 -\- a2 - I = 0, where ai and a2 are complex numbers implies that ai = a2 = 0). Similarly, the rows of {si — A)~^B =
lA^l ^
are linearly independent over the field of complex numbers. Also,
smce 0 (iii) rank ^ — rank \B^AB\ — rank 1
1 0
-1 Si
(iv) rank [5;7 —A,5] = rank
: n, or 0 1
2 = n for Si = 0,i = 1,2, the
eigenvalues of A. If 5
in place of T 0
(i) WriOJ)
then 0 (see Example 2.4) with rank Wr{0,T) = I <2 = n, and 0
\/s , neither of which has rows that are 0 linearly independent over the complex numbers. Also,
(ii) e^^B •
(iii) rank'
and {sI-A)-^B
1 0
0 0
•
I <2 = n, and
(iv) rank [sil — A,B\ = rank
-1 Si
1 OJ
I <2 = n for Si = 0.
Based on any of the above tests, it is concluded that the system is not reachable.
C. Discrete-Time Systems The response of discrete-time systems was studied in Section 2.7 of Chapter 2. We consider systems described by equations of the form
241 CHAPTER 3 :
Controllability, Observability, and Special Forms
242 Linear Systems
x(k + 1) - A(k)x(k) + B(kMkl
k ^ ko,
(2.47)
^j^^j,^ ^^^^ ^ j^nxn^ 5(^) ^ ^nxm^ ^j^^j ^^^ ^^^^^ ^^^^ ^ ^m ^^^ defined for it ^ ko. The state x(fe) for k> kois given by yt-l
4 ^ ) = ^(k ko)x(ko) + ^
^(k i + l)B{i)u{i),
(2.48)
/=^,
where the state transition matrix ^{k, ko) is given by ^(k, ko) = A(k - 1) X A(k - 2)'' 'A(ko) for k > ko, and O(^o. ko) = L In the time-invariant case we have x{k+\)
= Ax{k) + Bu{k\
k ^ ko,
(2.49)
where A G /^'^X" and B E /?"><^. The state x(yfc) of (2.49) is given by (2.48) with ^{k, ko) = A^-^',
k ^ ko.
(2.50)
Let the state at time ko be XQ. For the state at some time ki > ko to assume the value xi, an input u must exist that satisfies xi = 3>(A:i, ^o)^o + ^
^(kh i + l)B(i)u(i),
(2.51)
Reachability and controllability are defined for discrete-time systems in a completely analogous fashion as in the continuous-time case. The mathematical development, however, involves summations instead of integrals and is easier to deal with. The time-varying case can be developed in a manner similar to the time-invariant case. For this reason, the discrete-time time-varying case will not be developed presently. Instead, we shall concentrate on the time-invariant case. Note that some of the results given below for the discrete-time time-invariant case have already been presented in Section 3.1, Introduction. Discrete-time time-invariant systems For the time-invariant system x(k + 1) == Ax(k) + Bu(k) given in (2.49), the elapsed time fei-/co is of central interest, and we therefore take ^0 = Oandfei = K. Recalling that ^{k, 0) = A^, we rewrite (2.51) as xi = A^xo + ^A^-^'^^^Bu{i) /=o
(2.52)
when ^ > 0, or XX =
A^XO + '^KUK,
where
%K = [B, AB,...,
A^-^B]
and
UK = [u^{K - \), u^(K - 2 ) , . . . , u^(0)f.
(2.53) (2.54) (2.55)
The definitions of reachable state xi, reachable sub space Rr, and a system being {completely state) reachable, or the pair (A, B) being reachable, are the same as in the continuous-time case (see Definition 2.9, and use integer K in place of real time T). Similarly, the definitions of controllable state XQ, controllable subspace Re, and
a system being (completely state) controllable, or the pair {A, B) being controllable are similar to the corresponding concepts given in Definition 2.10 for the case of continuous-time systems. To determine the finite input sequence for discrete-time systems that will accomplish a desired state transfer, if such a sequence exists, one does not have to define matrices comparable to the reachability Gramian Wr, as in the case for continuoustime systems. In particular, we have the following result. THEOREM 2.18. Consider the system x{k + V) = Ax{k) + Bu{k) given in (2.39) and let x(0) = 0. There exists input u that transfers the state to xi in finite time if and only if xx E ^{%). In this case, xi is reachable and Rr = S^(^). An appropriate input sequence {u{k)}, k = 0,..., w - 1, that accompHshes this transfer in n steps is determined by Un = [u^(n l),u^(n - 2),..., w^(0)]^, a solufion to the equation "^Un = xi.
(2.56)
Henceforth, with an abuse of language, we will refer to Un as a control sequence when, in fact, we actually have in mind {u(k)}. Proof, In view of (2.52), xi can be reached from the origin in K steps if and only if xi = ^KUK has a solution UK, or if and only if x\ E ^(^K)- Furthermore, all input sequences that accomplish this are solutions to the equation x\ = ^RUK- For x\ to be reachable we must have xi E ^(^K) for some finite K. This range, however, cannot increase beyond the range of ^„ = % i.e., ^("^K) = S^(^n) for K > n. This follows from the Cayley-Hamilton Theorem, which implies that any vector x in ^(^A:), K > n, can be expressed as a linear combination of B, AB,..., A^~^B. Therefore, x E S?l(^„). It is of course possible to have x\ E ^{%K) with ^ < n for a particular x\\ however, in this case xi E S^(^„) since %K is a subset of ^„. Thus, xi is reachable if and only if it is in the range of ^„ = %. Clearly, any t/„ that accomplishes the transfer satisfies (2.56). • As pointed out in the above proof, for given x\ we may have x\ E ^(^K) for some K < n.\n this case the transfer can be accomplished in fewer than n steps, and appropriate inputs are obtained by solving the equation ^KUK = ^iCOROLLARY 2.19. The system x(k + 1) = Ax(k) + Bu(k) given in (2.49) is (completely state) reachable, or the pair (A, B) is reachable, if and only if rank'^ = n.
(2.51)
Proof, Apply Theorem 2.18, noting that S/l(^) = Rr = /?" if and only if rank ^ = n. THEOREM 2.20. There exists an input u that transfers the state of the system x(k + 1) = Ax(k) + Bu(k) given in (2.49) from XQ to x\ in some finite number of steps K, if and only if xi - A^xo E m.C^Kl Such input sequence UK = [u^(K -l),u^(K-2),..., the equation "^KUK = xi-A^xo. Proof The proof follows directly from (2.53).
(2.58)
u^(0)]^ is determined by solving (2.59) •
243 CHAPTER 3 :
Controllability, Observability, and Special Forms
244 Linear Systems
The above theorem leads to the following result that establishes the importance of reachability in determining u to transfer the state from any XQ to any xi in a finite number of steps. COROLLARY2.21. Let the sy Stem x(^+1) = Ax(^) + 5w(^) given in (2.49) be (completely state) reachable, or the pair (A, B) be reachable. Then there exists an input sequence to transfer the state from any XQ to any x\ in a finite number of steps. Such input is determined by solving Eq. (2.60). Proof, Consider (2.54). Since (A, B) is reachable, r<2n^^„ = rank% = nandS/l(^) = R^. Then %Un = XX - A " x o
(2.60)
always has a solution Un = [u^(n - 1 ) , . . . , w^(0)]^ for any XQ and xi. This input sequence transfers the state from XQ to xi in n steps. • Note that, in view of Theorem 2.20, for a particular XQ and xi, the state transfer may be accomplished in K < n steps, using (2.59). EXAMPLE 2.8. Consider the system in Example 1.1, namely, x(k + 1) = Ax(k) + 0 1 0 1 5M(^),whereA = .SincQ rank^ = rank[B,AB] = rank and 5 = 1 1 1 1 2 = n, the system is reachable, and any state XQ can be transferred to any other state 0 1 u(l) xi in 2 steps. Let xi = . Then (2.60) implies that ,Xo = 1 1 u(0) u(l) -1 1 1 1 0 1 ao b-l-bo This agrees or u(0) 1 2 1 0 1 1 bo} a- ao - bo with the results obtained in Example 1.1. In view of (2.59), if Xi and XQ are chosen so that 0 1 a- bo is in the 2/1(^0 - ^(B) = span xi - Axo 1 1 b- ao- bo. then the state transfer can be achieved in one step. For example, if xi = Xo
thenBu(0)
=
u(0) = xi - Axo =
and
implies that the transfer from xo to
xi can be accomplished in this case in 1 < 2 = fz steps with u(0) = 2. EXAMPLE 2.9. Consider the system x(k + 1) = Ax(k) + Bu(k) with A
0 0
1 and 0
0 1 has full rank, there exists an input sequence that 1 0 will transfer the state from any x{0) = xo to any x(n) = xi (in n steps), given by (2.60), U{1) 0 1 (xi - Xo). Compare this with Example 2.5, \xi - A^xo) = U2 = Lw(0)J 1 0 where the continuous-time system had the same system parameters A and B. • B =
. Since ^ = [B, AB]
We shall now establish the relationship between reachability and controllability for the discrete-time time-invariant systems x(k + 1) = Ax(k) + Bu(k) given in (2.49). Consider (2.51). The state XQ is controllable if it can be steered to the origin xi = 0 in a finite number of steps K. That is, XQ is controllable if and only if A^xo
=
%KUK
(2.61)
245
for some K, or when
(2.62) for some K. Recall that x\ is reachable when
(2.63) THEOREM 2.22. Consider the system x{k + 1) = Ax{k) + Bu{k) given in (2.49). (i) If state X is reachable, then it is controllable, (ii) RrdRc(iii) If the system is (completely state) reachable, or the pair (A, B) is reachable, then the system is also (completely state) controllable, or the pair (A, B) is controllable. Furthermore, if A is nonsingular, then relations (i) and (iii) become if and only if statements, since controllability also implies reachability, and relation (ii) becomes an equality, i.e., Re = Rr. Proof, (i) If X is reachable, then x G S/l(^). In view of Lemma 2.14, S/l(^) is an Ainvariant subspace and so A"x G S/l(^), which in view of (2.61) implies that x is also controllable. Since x is an arbitrary vector in Rr, this implies (ii). If S^(^) = /?", the whole state space, then A^x for any x is in S/l(^) and so any vector x is also controllable. Thus, reachability implies controllability. Now, if A is nonsingular, then A~" exists. If x is controllable, i.e., A^x G 2?l(^), then x G ^ ( ^ ) , i.e., x is also reachable. This can be seen by noting that A~" can be written as a power series in terms of A, which in view of Lemma 2.15, implies that A~"(A"jc) = ;c is also in ^(%). • Matrix A being nonsingular is the necessary and sufficient condition for the state transition matrix 0(/:, ^o) to be nonsingular (see Section 2.8), which in turn is the condition for "time reversibility'' in discrete-time systems. Recall that reversibility in time may not be present in such systems since 0(^, ^o) may be singular. In contrast to this, in continuous-time systems, ^(t, to) is always nonsingular. This causes differences in behavior between continuous- and discrete-time systems and implies that in discrete-time systems controllability may not imply reachability (see Theorem 2.22). Note that, in view of Theorem 2.16, in the case of continuous-time systems, it is not only reachability which always implies controllability, but also vice versa, controllability always implies reachability. In the following, we introduce the discrete-time reachability and controllability Gramians for system (2.49). These are defined in a manner analogous to the continuous-time case. DEFINITION2.13. The reachability Gramian is defined by K-l
Wr(0, K) = ^ A^-^^-'^^BB^(A^f-^'^^\ /=o It is not difficult to verify that W,(0, K) = Xf~o A'BB^{A^j = "^K^-
(2.64) •
LEMMA 2.23. ^{%) = ^(Wr(0, K)) for every K ^ n. Proof, This result can be established in a way similar to the proof of the corresponding result in the continuous-time case (Lemma 2.7). The details are left to the reader. • When a system is reachable, the input sequence that transfers XQ aX k = 0 to x\a.tk = K can be determined in terms of the reachability Gramian. In particular, let
CHAPTER 3 :
Controllability, Observability, and Special Forms
246
rank % = n = rank Wr{^, K) for K > n, and notice that the relation
Linear Systems
UK = '^KW;\0,
K)(xi - A^xo)
(2.65)
satisfies (2.59) since W,(0, K) = ^ / ^ ^ J . DEFINITION2.14. The controllability Gramian is defined as Wc{^, K) = ^
A'^'^^^BB^(A^y^'^^\
(2.66)
We note that WdO, K) is well defined only when A is nonsingular. The reachability and controllability Gramians are related by Wr{0,K) =
A^WMK)(A^f,
(2.67)
as can easily be verified. When A is nonsingular, the input that will transfer the state from XQ at /: = 0 to xi = Oinn steps can be determined using (2.60). In particular, one needs to solve [A-^'^Wn = [A-''B,...,A-^B]Un = xo
(2.68)
for Un = [u^(n-1),..., w^(0)]^. Note that XQ is controllable if and only if -A^XQ G g/l(^), or if and only if XQ G g i ( A ~ " ^ ) for A nonsingular. Clearly, in the case of controllability (and under the assumption that A is nonsingular), the matrix A~^% is of interest, instead of ^ [see also (1.18)]. In particular, a system is controllable if and only if ranfc (A~'^^) = rank% = n. EXAMPLE 2.10. Consider the system x(k + 1) = Ax(k) + Bu(k), where A =
1 1 0 1
1 1 = \ <2 = n, this system 0 0 is not (completely) reachable (controllable-from-the-origin). All reachable states are of and 5
the form a
. Since ra^^^ = rank[B, AB] = rank
, where a E R since
is a basis for the 2?l(^) = Rf, the reachability
subspace. The reachability Gramian for i^ = n = 2isWX0, 2) = BB^ + (AB)(AB)T _ 1 0" 1 0 '2 0' + Note that a basis for ^(Wr(0, 2)), is and S?l(^) = 0 0 0 0 0 0 ^(Wr(0, 2)), which verifies Lemma 2.23. In view of (2.62) and the Cay ley-Hamilton Theorem, all controllable states XQ satisfy A^xo E 9l(^); i.e., all controllable states are of the form <^ L , where a G R. This verifies Theorem 2.22 for the case when A is nonsingular. Note that presently Rr = Re EXAMPLE 2.11. Consider the system x{k + 1) = Ax(k) + Bu{k), where A and 5
SincQ rank ^ = rank[B,AB]
= rank
1 0 = 1 < 2 == ^, the system 0 0
is not (completely) reachable. All reachable states are of the form a smce
0 1 0 0
is a basis for ^(^) = Ry, the reachability subspace.
where a E: R
To determine the controllable subspace R^ consider (2.62) for K = n/m view of the Cay ley-Hamilton Theorem. Note that A~^% cannot be used in the present case, since A is singular. Since A^XQ =
G S/l(^), any state XQ will be a controllable
Xo
state, i.e., the system is (completely) controllable and Re = R^. This verifies Theorem 2.22 and illustrates that controllability does not in general imply reachability. Note that (2.60) can be used to determine the control sequence that will drive any state xo to the origin (xi = 0). In particular, u(l) u(0)
^f/„ =
=
-A^XQ.
Therefore, w(0) - a and w(l) = 0, where a E /? will drive any state to the origin. To verify this, we consider x(l) x(2) = Ax(l) + Bu(l) --
Ax(0) + Bu(0) X02 + Oi
0
0
Xoi
X02 + ex
0 <
•^02.
0
and
0 =
3.3 OBSERVABILITY AND CONSTRUCTIBILITY In applications, the state of a system is frequently required but not accessible. Under such conditions, the question arises whether it is possible to determine the state by observing the response of a system to some input over some period of time. It turns out that the answer to this question is affirmative if the system is observable. Observability refers to the ability of determining the present state x(to) from knowledge of future system outputs, y(t), and system inputs, u(t), t ^ to. Constructibility refers to the ability of determining the present state x(fo) from knowledge of past system outputs, y{t), and system inputs, u{t), t ^ to. Observability was briefly addressed in Section 3.1. In this section this concept is formally defined and the (present) state is explicitly determined from input and output measurements. As in Section 3.2 (dealing with reachability and controllability), the reader can concentrate on the time-invariant case, starting with Definition 3.9, if so desired, and omit the more general material (dealing with time-varying systems) that starts with Definition 3.1. Section description In this section, observability and constructibility are introduced and discussed in detail for given linear system state-space descriptions. This is accomplished for continuous and discrete-time systems and for both time-varying and time-invariant cases. Observability in continuous-time systems is addressed first with introduction of the observability Gramian Wo(to, ti) (Definition 3.4). It is shown (Corollary 3.2) that a system is observable at to if Wo(to, ti) has full rank for some ti > to, and furthermore, if the system is observable, how an initial state can be determined. Observability refers to the ability to determine the current state of a system from future system outputs (and inputs). Constructibility, which refers to the ability of determining the current state of a system from past system outputs (and inputs), is addressed next with the introduction of the constructibility Gramian (Definition 3.8). It is shown
247 CHAPTERS:
Controllability, Observability, and Special Forms
248 Linear Systems
(Theorem 3.3) that a continuous-time system is observable if and only if it is constructible. Next, additional tests for observability and constmctibility are obtained (Theorem 3.4). All these results are then applied in the study of the continuous-time time-invariant case. The material is arranged in such a manner that the continuoustime time-invariant systems can be studied independently of the time-varying case. The relation between the observability Gramian Wo(0, T) and the observability matrix 0 = [C^, {CAf,..., (CA""^)^]^ is estabhshed next (Lemma 3.6). It is then shown (Corollary 3.8) that a system is observable if and only if 0 has full rank n. Next, the relation between observability and constmctibility is established (Theorem 3.9). Finally, additional tests for observability are derived (Theorem 3.10). A discussion of observability and constmctibility for discrete-time systems with particular emphasis on time-invariant systems is presented next. It is shown that a system is observable if the observability matrix 0 has full rank (Corollary 3.12), and for this case, an expression for the initial state XQ is given as a function of future outputs (and inputs). A similar result involving the observability Gramian Wo{0, K) in place of 0 is also established (Corollary 3.14). Next, it is shown that observability in discrete-time systems always implies constmctibility. In contrast to the continuous-time case, the converse of the above statement is generally not true. This is due to the lack of time reversibility in difference equations. When A is nonsingular, then constmctibility also implies observability. Finally, the constmctibility Gramian Wcnd^y K) is also introduced (Definition 3.16). A. Continuous-Time Time-Varying Systems We consider systems described by equations of the form X = A(t)x + B(t)u,
y = C{t)x -H D{t)u,
(3.1)
where A{t) G iR^^^ B{t) G R''''^, C{t) E 7?^><^ D{t) G /?^><^, and u{t) G R^ are defined and (piecewise) continuous on some real open interval {a, b). It was shown in Chapter 2 that the output y{t) is given by t
y{t) = C(t)^(t, to)x(to) +
C(t)
(3.2)
Jto
for to, t G {a, h), where ^{t, r) denotes the state transition matrix. This can be written as y{t) - C{tmu where j ( 0 = y(t)
to)xo,
^ C(t)(t, r)B(r)u(T)dT
(3.3)
+ D{t)u(t)andxo = x(ro)-Wewill
find it convenient to first give the definition of an unobservable state. DEFINITIONS.1. A State X is unobservable at time to if the zero-input response of the system is zero for every t > to, i.e., if C(t)(^(t, to)x = 0
for every t > to.
(3.4)
• DEFINITION 3.2. The unobservable at to subspace R^^ of (3.1) is R^^ = { set of all unobservable at ^o states x}.
•
When the context is clear and there is no ambiguity, we will write Ro in place ofR'l
DEFINITION 3.3. The system (3.1) is {completely state) observable at ^ , or the pair (A(0, C{t)) is observable at ^ , if the only state x G /?" that is unobservable at fo is the zero state, jc = 0, i.e., if R^^ = {0}. • We will show later that observability depends only on the pair (A(t), C(t)). According to Definition 3.1, a nonzero unobservable state x cannot be distinguished from the zero state if only the (future) outputs are known; that is, an unobservable state cannot be determined uniquely from knowledge of the inputs and outputs of the system. This can be seen from (3.3), where for the unobservable states x at ^0. y(t) = 0 for r ^ to. This implies that the states x, which together with the input u(t) produce the output y(t), cannot be distinguished from the zero state, since they both produce the same output. In a manner analogous to the development of reachability in Section 3.2, we make the following definition. DEFINITION 3.4. The observability Gramian of the system (3.1) is the n X ri matrix Wo{to,tx)=
(3.5)
Note that Wo is symmetric and positive semidefinite for every ti > to, i.e., Wo = W j and Wo^ 0 (show this). THEOREM 3.1. A State X is unobservable at to if and only if (3.6)
xGX(Wo(to.ti)) for every t\ > ^o, where J{(-) denotes the null space of a map.
Proof. If X is unobservable, then (3.4) is satisfied. Postmultiply (3.5) by x to obtain Wo(to, ti)x = 0 for every ti > to, i.e., x G }((Wo(to, ti)) for every ^i > to. Conversely, let X be in the null space of Wo. Then x'^Wo(to, ti)x = |^^^ | | C ( T ) ^ ( T , ro)x|p dr = 0 for every ti > to. This implies that (3.4) is true, or that the state x is unobservable. • -1 e^^ 0 -1 and C(t) C(0 = [0, e~^]. The state transition matrix in this case is (see Example 2.1 in this chapter)
EXAMPLE 3.1. Consider the system i: = A(t)x, y = C(0^, where A(0 =
^-(t-r)
0(r, r) = Then C(T)^(T,
i^^t+T
_ ^-r+Sr-)
e
0
to) = [0, e ^^+^0] and the observability Gramian is given by
WoitoJi) =
0
^2^0
,-2T
to
[0,e~^']dT
0
0
0
^-4^
dr
0
0
0
^-4/1 _
^-4to
It is clear that this system is not observable, since rank Woito, ti) = I <2 = n.ln view of Theorem 3.1, all unobservable states are given by Notice that };(0 = C(t)^(t,to)xo of the (unobservable) states
0
, where a E R.
= [0, e-^^^'o]xo = 0 for JCQ =
can be distinguished from the zero state.
, that is, none
249 CHAPTER 3 :
Controllability, Observability, and Special Forms
250 Linear Systems
Clearly, x is observable at t^ if and only if there exists a fi > ^o such that ^ovo, ti)X ¥" 0. COROLLARY 3.2. The system (3.1) is {completely state) observable at to, or the pair (A(t), C(t)) is observable at to, if and only if there exists afiniteti > to such that rankWoitoJi) = n.
(3.7)
If the system is observable, the state xo at to is given by fh Xo =
W^\to,ti)
CD^(T,^O)C^(T)KT)^T
(3.8)
^0
Proof, The system is observable if and only if the only vector x that satisfies (3.6) is the zero vector. This is true if and only if there exists (at least one)finitetime ti for which the null space of Wo(to, ti) contains only the zero vector, or if and only if (3.7) is satisfied. To determine the state XQ at to, given the output and input values over [^o. ^i], premultiply (3.3) by (^^(t, to)C'^(t) and integrate over [to, ti]. Then, in view of (3.5), Woito, h)xo = r
CD^(T, fo)C^(T)KT) Jr.
When the system is observable, (3.9) has the unique solution (3.8).
(3.9) •
It is clear that if the state at some time to is found, then the state x(t) for r ^ ^0 is easily determined, given u(t), t > to, via the variation of constants formula (2.2). We mention here that alternative methods to (3.8) to determine the state of a system when the system is observable are given in the next chapter (in Section 4.3 on state estimation). EXAMPLE 3.2 For the scalar system x = a{t)x,y = c{t)x, where a{t) = - 1 and c{t) = e\ we have and C{T)^(J, to) = e^e'^'"'^^^ = e^^. The observability Gramian in this case is Woito, ^i) = //^ e'^^^ dr = e'^^^it\ - to), which implies that the system is observable at ^o since rankWoito, t\) = 1 = w for any t\ y^ to. Suppose now that the observed output is y(t) = y(t) = ae^^ fort > to. Then xo can be determined using (3.8), Xo = [e-^'^/(ti - to)][\,l'e\ae')dT] = a. Indeed, y(t) = c(t)^(t, to)a = ae^o, as observed. • Observability utilizes future output measurements to determine the present state. In contructibility, past output measurements are used to accomplish this. Constructibility is defined below and its relation to observability is established. DEFINITION 3.5. A State X is unconstructible at time t\ if for everyfinitetime t ^ t\, the zero-input response of the system is zero for all t, i.e., C{t)^{t, ti)x = 0
for every t < ^i.
(3.10)
DEFINITION3.6. The unconstructible at t\ sub space S^^ of (3.1) is (3^^ = {set of all states x unconstructible at ^i}.
•
It is denoted in the following by R^, for convenience, when there is no ambiguity.
DEFINITION3.7. The system (3.1) is {completely state) constructible at t\, or the pair (A(0, C{t)) is constructible at ti, if the only state x G R" that is unconstructible at to is X = 0,i.e.,if/?^ = {0}. • THEOREM 3.3. If the system (3.1), or the pair (A(0, C(0), is observable at ^ , then it is constructible at some ti > ^ ; if it is constructible at ti, then it is observable at some to^ ti. It was shown in Corollary 3.2 that the system is observable at ^o if and only if rank Wo(to, t\) = n for some ti > to. Similar results can be established for constructibility. We will require the following concept. DEFINITION 3.8. The constructibUity Gramian of (3.1) is the n X n matrix ^ ^ ( r , ti)C^{T)C{T)^{T,
Wcn{to,tl)
tOdr.
(3.11)
Proof of Theorem 3.3. A similar result as Theorem 3.1, but for unconstructible state x, can be derived. Next, using a proof similar to the proof of Corollary 3.2, it can be shown that the system is constructible at ti if and only if there exists finite to < ti such that rank Wcn(to> ti) = n.
(3.12)
=
(3.13)
Note that WoitoJl)
^\tiJoWcn(to>tO<^(thtol
which implies that rank Wo(to, ti) = rank Wcn(to, h) for every to and ^i. Note that ^(ti, to) is nonsingular for every to and ti. • EXAMPLE 3.3. (i) Consider the system x = A{t)x, y constructibility Gramian is 0
Wcnito, h) =
[0,e
-2T + t]
0 0
C{t)x of Example 3.1. The
]dT
4
Compare this with the observabiUty Gramian of Example 3.1 and note that ^HhJoWcn(to,h)^ituto)
=
0
-\e^''
Wo(to,til
as expected [see (3.13)]. Presently, C{t)^{t,ti)x
= [0,e
-2t+t^i
0
for every t ^ ti implies, in view of Definition 3.5, that all unreconstructible states (at ^i) are of the form , a G /?. Note that they are identical to the unobservable states (see Example 3.1). (ii) For the scalar system x = -x,y = e^ of Example 3.2, the constructibility Gramian is Wcn(to, h) = e^^^iti - to) and ^^(^i, to)Wcn(to, ^1)^(^1, to) = ^-(^1-^0)^2^1 >< (ti - ro)e~(^i~^o) = e^^^iti - ^o) = Wo(to, ti), as expected in view of (3.13). •
We shall now use Lemma 2.7 in Section 3.2 to develop additional tests for observability and constructibility. These are analogous to corresponding results that we established for reachability and controllability (Theorems 2.8 and 2.9).
251 CHAPTER 3 :
Controllability, Observability, and Special Forms
252 Lh^^Systems
THEOREM 3.4. The system i = A(t)x+B(t)u, y = C(0-^+/)(Ow is (completely state) observable at ^o (i) if and only if there exists afiniteti > to such that rankWo(to,ti) = n,
(3.14)
where Wo{tQ,t\) = \^^^^ <^^(T,to)C'^(T)C(T)^(TJo)dT is the observability Gramian, or equivalently, (ii) if and only if there exists finite ti > to such that the n columns of C(t)^(tJo)
(3.15)
are linearly independent on [to, ti] over the field of complex numbers. Proof. Part (i) was shown in Corollary 3.2 and part (ii) is a consequence of Lemma 2.7 (compare with the corresponding Theorem 2.8 for reachability). • Similar results can be derived for constructibility. In particular, we have the following result. THEOREM 3.5. The system x = A(t)x + B(t)u, y = C(t)x + D(t)u is (completely state) constructible at ti (i) if and only if there exists finite to < t\ such that rankWcn{to,t\) = n,
(3.16)
where WcnitoJi) = ^^^^ ^^(rJi)C'^(r)C(r)^(T,ti)dT, the constructibility Gramian, or equivalently, (ii) if and only if there exists finite to < ti, such that the n columns of C(tmt,ti)
(3.17)
are linearly independent on [to, ti] over the field of complex numbers. Proof, The proof is analogous to the proof of the corresponding results on observability. • Note that postmultiplication of C(0^(^, ^i) by the nonsingular matrix ^(ti, to) yields C(t)^(t, to) in (ii) of Theorem 3.4 [compare with (3.13)]. This shows again the result given in Theorem 3.3 that observability implies and is implied by constructibility, in the case of continuous-time systems.
B. Continuous-Time Time-Invariant Systems We shall now study observability and constructibility for time-invariant systems described by equations of the form x = Ax-^Bu,
y = Cx + Du,
(3.18)
where A G /^^^^^ B E T^^^^, C G /?^><^ D G 7^^><^, and u(t) G R"^ is (piecewise) continuous. As was shown in Chapter 2, the output of this system is given by rt
Ce^^'-'^Bu{7)dT
y(t) = Ce'^'xiO) +
+ Du{t).
(3.19)
0
We recall once more that in the present case (r, r) = ^(t - r, 0) = exp [A{t - T)] and that initial time can always be taken to be to = 0. We will find
253
it convenient to rewrite (3.19) as
m whcvcy{t)
= y{t)-
:C/%,
\J^Ce^^'-^^Bu{T)dT^Du{t) andxo
(3.20) =x(0).
DEFINITION 3.9. A state x is unobservable if the zero-input response of the system (3.18) is zero for every ^ > 0, i.e., if -- 0
for every ^ > 0.
(3.21)
The set of all unobservable states, x,R^, is called the unobservable subspace of (3.18). System (3.18) is (completely state) observable, or the pair (A,C) is observable, if the only state x e R^ that is unobservable is x = 0, i.e., if R^ = {0}. Definition 3.9 states that a state is unobservable precisely when it cannot be distinguished as an initial condition at time 0 from the initial condition x(0) = 0. This is because in this case the output is the same as if the initial condition were the zero vector. DEFINITION 3.10. The observability Gramian of system (3.18) is \hQnxn Wo(0, T)^ Vo{()J)=
^C Ce^^dT. I[ e""^ e^^'C^Ce^'dT. Jo
matrix (3.22)
We note that Wo is symmetric and positive semidefinite for every T > 0, i.e., ^o = ^J and Wo>0 (show this). Recall that the pnxn observability matrix C CA ^:
(3.23) CA n-\
was defined in Section 3.1. We now show that the null space of ^ ^ ( 0 , 7 ) , denoted by ^ ( ^ ^ ( 0 , 7 ) ) , is independent of T, i.e., it is the same for any T > 0, and in particular, it is equal to the null space of the observability matrix 0'. Thus, the unobservable subspace RQ of the system is given by the null space of ^ , ^ ( ^ ) , or the null space of Wo(0, T), ^ ( W o ( 0 , r ) ) for some finite (and therefore for all) T > 0. LEMMA3.6. J/{0)
= ^(Wo(0,T))
for every 7 > 0.
Proof, If X G ^ ( ^ ) , then ^x = 0. Thus, CA^x = 0 for all 0 < ^ < n - 1, which is also true for every ^ > n — 1, in view of the Cay ley-Hamilton Theorem. Then Ce^^x = C[E^=o(^^/^-Mi-^ = 0 for every finite t. Therefore, in view of (3.22) ^^^(0, T)x = 0 for every 7 > 0, i.e., x G yK{Wo{0, T)) for every 7 > 0. Now letx G yK{Wo{0, T)) for some r > 0, so that x^ W{0, T)x = f^^0 II Ce^^x f dT = 0, or Ce^'x = 0 for every t G [0, T]. Taking derivatives of the last equation with respect to t and evaluating at ^ = 0, we obtain Cx = CAx = • • • = CA^x = 0 for every ^ > 0. Therefore, CA^x = 0 for every ^ > 0 , i.e., ^ x = O o r x G ^ ( ^ ) . • THEOREM 3.7. A State X is unobservable if and only if XG^(^),
(3.24)
xe^(Wo{0,T))
(3.25)
or equivalently, if and only if
CHAPTER 3 :
Controllability, Observability, and Special Forms
254 Lh^Systems
for some finite (and therefore for all) T > 0. Thus, the unobservable subspace Ro = ^ ( ^ ) = ^ ( ^ o ( 0 , T)) for some T >0. Proof, If X is unobservable, (3.21) is satisfied. Taking derivatives with respect to t and evaluating at ^ = 0, we obtain Cx = CAx = ••• = CA^x = 0 for ^ > 0 or CA^x = 0 for every ^ > 0. Therefore, ^x = 0 and (3.24) is satisfied. Assume now that ^x = 0, i.e., CA^x = O f o r O < ^ < n — 1 , which is also true for every ^ > n — 1, in view of the Cay leyHamilton Theorem. Then Ce^^x = C[Y^^Q{t^/k\)A']x = 0 for every finite t, i.e., (3.21) is satisfied and x is unobservable. Therefore, x is unobservable if and only if (3.24) is satisfied. In view of Lemma 3.6, (3.25) follows. • Clearly, x is observable if and only if ^ x 7^ 0 or ^ ^ ( 0 , T)x ^ 0 for some T > 0. COROLLARY 3.8. The system (3.19) is (completely state) observable, or the pair (A, C) is observable, if and only if rank ^ = n,
(3.26)
rankWo{0,T)=n
(3.27)
or equivalently, if and only if
for some finite (and therefore for all) 7 > 0. If the system is observable, the state XQ at ^ = 0 is given by
xo =
W-\OJ)
(3.28)
Jo
Proof, The system is observable if and only if the only vector that satisfies (3.20) or (3.21) is the zero vector. This is true if and only if the null space is empty, i.e., if and only if (3.26) or (3.27) are true. To determine the state XQ at ^ = 0, given the output and input values over some interval [0,7], we premultiply (3.20) by e^ ^C^ and integrate over [0, T] to obtain Wo{0,T)xo=
[ /''C^y{T)dT, (3.29) Jo in view of (3.22). When the system is observable, (3.29) has the unique solution (3.28).
• Note that T > 0, the time span over which the input and output are observed, is arbitrary. Intuitively, one would expect in practice to have difficulties in evaluating XQ accurately when T is small, using any numerical method. Note that for very small r , |Wo(0,r)| can be very small, which can lead to numerical difficulties in solving (3.29). Compare this with the analogous case for reachability, where small T leads in general to large values in control action. It is clear that if the state at some time t^ is determined, then the state x{t) at any subsequent time is easily determined, given w(f), f > fo, via the variation of constants formula (3.2), where 0(f, T) = exp[A(f - T ) ] . Alternative methods to (3.29) to determine the state of the system when the system is observable are provided in the next chapter, in Section 4.3. EXAMPLE 3.4. (i) Consider the system x = Ax,y = Cx, where A C = [ 1 , 0 ] . Here e^'
1 0
0 0
1 and 0
t and Ce^^ = [l,t]. The observability Gramian is then 1
Wo(0,T) = \n
\[lT]dT = \^
1
T
T
T
\rp2
255 NoticQih2itdetWo(0,T)
dr 3^
j^r^ 7^ 0 for any T > 0, i.e., ra^^ W^(0, T) = 2 = fz for any T > 0, and therefore (Corollary 3.8), the system is observable. Alternatively, note that the observabilC 1 0 and rank€ = 2 = n. Clearly, in this case J{(€) ity matrix 0 = 0 1 CA 0 , which verifies Lemma 3.6. M(Wo(0, T)) = 0 0 1 , as before, but C = [0, 1], in place of [1, 0], then Ce^^ = [0, 1] (ii)If A = 0 0 0 0 and the observability Gramian is WoiO, T) = [0,1] J T We have 0 T rank Wo(0,T) = I < 2 = n and the system is not completely observable. In view of Theorem 3.7, all unobservable states x E XiWoiO, T)) and are therefore of the form 0 1 Note that ,a E^ R. Alternatively, the observability matrix 0 = c 0 0 CA
jvr(O) = >r(Wo(o, T))
span
Observability utilizes future output measurements to determine the present state. In (re)constructibility, past output measurements are used. Constructibility is defined in the following, and its relation to observability is determined. DEFlNITlON3.il. A State x is unconstructible if the zero-input response of the system (3.18) is zero for all t < 0, i.e., Ce^'x = 0
for every ^ < 0.
(3.30)
The set of all unconstructible states x, R^, is called the unconstructible subspace of (3.18). The system (3.18) is (completely state) {re)constructible, or the pair (A, C) is (re)constructible, if the only state x E /?" that is unconstructible is ;c = 0, i.e., Ren = {0}. We shall now establish a relationship between observability and constructibility for the continuous-time time-invariant systems (3.18). Recall that x is unobservable if and only if Ce'^'x = 0
for every t > 0.
(3.31)
THEOREM 3.9. Consider the system x = Ax + Bu, y = Cx + Du given in (3.18). (i) A state x is unobservable if and only if it is unconstructible. (ii) Ro = R^. (iii) The system, or the pair (A, C), is (completely state) observable if and only if it is (completely state) (re)constructible. Proof, (i) If X is unobservable, then Ce^^x = 0 for every t > 0. Taking derivatives with respect to t and evaluating ai t = 0, we obtain Cx = CAx = "• = CA^x = 0 for A: > 0 or CA^'x = 0 for every /: > 0. This, in view of Ce^^x = ^"l=Q(t^/k\)CA^x, implies thai Ce^^x = 0 for every r < 0, i.e., x is unconstructible. The converse is proved in a similar manner. Parts (ii) and (iii) of the theorem follow directly from (i). • The observability Gramian for the time-invariant case, Wo(0, T), was defined in (3.22). In view of (3.11), we make the following definition.
CHAPTER 3 :
Controllability, Observability, and Special Forms
256
DEFINITION 3.12. The constructibUity Gramian of system (3.18) is ih^nXn
Linear Systems
,AHr~T)^T^^A(r-T)^^^
W,,(0, T)
matrix (3.32)
Note that Wo(0,T)
=
e^"^Wcn{0,T)e AT
(3.33)
as can be verified directly [see also (3.13)]. As in the time-varying case above, we now introduce a number of additional criteria for observability. THEOREM 3.10. The system x = Ax + Bu, y = Cx + Dui^ observable (i) if and only if rank W^(0, T) = n
(3.34)
for some finite T > 0, where Wo(0, T) = \Q e^^'^C^Ce^' dr, the observability Gramian, or (ii) if and only if the n columns of Ce"^'
(3.35)
are linearly independent on [0, oo) over the field of complex numbers, or alternatively, if and only if the n columns of C(sl - A)-^
(3.36)
are linearly independent over the field of complex numbers, or (iii) if and only if rank € = n,
where 0 -
C CA
(3.37)
the observability matrix, or
CA n - l (iv) if and only if rank
Sil - A
C
(3.38)
for all complex numbers 5/, or alternatively, for all eigenvalues of A. Proof. The proof of this theorem is completely analogous to the (dual) results on reachability (Theorem 2.17) and is omitted. • Similar results as those given in Theorem 3.10 can be derived for constructibility, and the reader is encouraged to state and prove these for the cases (i) and (ii). This is of course not surprising, since it was shown (in Theorem 3.9) that observability implies and is implied by constructibility. Accordingly, the tests developed in the theorem for observability are typically also used to test for constructibility. EXAMPLE 3.5. Consider the system x = Ax, y = Cx, where A =
andC 0 0 [1, 0], as in Example 3.4(i). We shall verify (i) to (iv) of Theorem 3.10 for this case.
(i) For the observability Gramian, Wo (0, r ) =
2^
iT2 2^
we have rank Wo(0, T)
= 2 = nfor any T > 0. (ii) The columns of Ce^^ = [l,t] are linearly independent on [0, oo) over the field of complex numbers, since ai • I + a2-1 = 0 implies that the complex numbers ai and a2 must both be zero. Similarly, the columns of C{sl — A)~^ = [1/5, \ls^] are linearly independent over thefieldof complex numbers. "1 0' (iii) rank€} = rank r c' = rank = 2 = n. .0 1. [cA_ (iv) rank
Sil - A]
C J
'Si
- 1 "
0
Si
= rank
.1
= = 2 = n for Si = 0, / = 1, 2, the eigen-
0.
values of A. Consider again A = 0 1 butC = [0, 1] [in place of [1,0], as in Example 3.4(ii)]. 0 0 The system is not observable for the reasons given below. (i) Wo(0, T) =
with rank Wo{0, T) = l< 2 = n.
(ii) Ce^^ = [0,1] and its columns are not linearly independent. Similarly, the columns of C(sl — A)~^ = [0, l/s] are not linearly independent. 0 1 C (iii) rank€ = rank 1<2 = n. = rank CA 0 0 -1 Sil - A = I < 2 = n for Si = 0, an eigenvalue = rank (iv) rank C of A.
C. Discrete-Time Systems We consider systems described by equations of the form x(k + 1) = A(k)x(k) + B(k)u(k\
y(k) = C(k)x(k) + D(k)u(k\
(3.39)
where A(k) G /('"^^ B(k) G /^"^"^, C(k) G i?^><^ D(k) G RP"""^, and the input u(k) G R"^ are defined for k ^ ko (see Section 2.7). The output y(k) for ^ > ^Q is given by k-\
y(k) = C(k)^(k
ko)x(ko) + ^
C(k)^(k, i + l)B(i)u(i) + D(k)u(kl
(3.40)
i = ko
where the state transition matrix ^(k, ko) is given by ^(k, ko) = A(k — l)A(k — 2) . . . A(ko) for k > ko, and
y(k) = Cx(k) + Du(k),
k ^ ko,
(3.41)
where A G iR"><^ C G jR^^^, C G /^^x^ D G RP"""^, and (3.40) is still valid with the state transition matrix 0(fc, ko) given in this case by ^(k, ko) = A^-^o,
k ^ ko.
(3.42)
257 CHAPTERS:
Controllability, Observability, and Special Forms
258 Linear Systems
Observability and (re)constmctibility for discrete-time systems are defined as ^^ the continuous-time case. Observability refers to the ability to uniquely determine the state from knowledge of current and future outputs and inputs, while constructibility refers to the ability to determine the state from knowledge of current and past outputs and inputs. In discrete-time systems, the time-varying case can be developed in a manner analogous to the time-invariant case and will therefore not be developed here. Instead, we shall concentrate on the time-invariant case. Note that some of the following results have already been presented in Section 3.1. Discrete-time time-invariant systems Consider the time-invariant system (3.41) and the expression for its output y(k), given in (3.40). Without loss of generality, we take ^o = 0. Then k-i
y(k) = CA^x{0) + ^ CA^-^'^^'^Buii) + Du{k) /=o
(3.43)
for yfe > 0 and j(0) = CJC(O) + Dw(0). Rewrite (3.43) as y{k) = CA^xo for ^ > 0, where y{k) = y(k)
(3.44)
XfZo CA^-^'+^^Bu(i) -h Du(k) for fc > 0 and
y(0) = y(0) - Du(Ol and XQ = x(0). DEFINITION 3.13. A State x is unobservable if the zero-input response of system (3.41) is zero for all ^ > 0, i.e., if CA^x = 0
for every y^ > 0.
(3.45)
The set of all unobservable states x, Ro, is called the unobservable subspace of (3.41). The system (3.41) is {completely state) obse^able, or the pair (A, C) is observable, if the only state x E /?" that is unobservable is x = 0, i.e., if Ro = {0}. • The pnX n observability matrix 0 was defined in (3.23). Let J{(G) denote the null space of 0. THEOREM3.il. A State X is unobservable if and only if X E K(€X
(3.46)
i.e., the unobservable subspace Ro = M(€). Proof. If X E Jvr(O), then Ox = 0 or CA^x = O f o r O < ^ < n - l . This statement is also true fork>n1, in view of the Cay ley-Hamilton Theorem. Therefore, (3.45) is satisfied and x is unobservable. Conversely, if x is unobservable, then (3.45) is satisfied and €x = 0. • Clearly, x is observable if and only if €x ¥" 0. COROLLARY 3.12. The system (3.41) is (completely state) observable, or the pair (A, C) is observable, if and only if rank€ = n.
(3.47)
If the system is observable, the state XQ at ^ = 0 can be determined as the unique solution of (3.48) where l'0.«-l = b ^ ( 0 ) , / ( l ) , . . . , / ( « - l ) ] ^ 6 / ? ' ' « , t/o.«-i = [M^(0),M^(l),...,«^(n-l)]^ €/?"•", and Mfi is the pn x mn matrix given by D CB
0 D
CA^'-^B
CA^'-^B
0 0
0" 0
D CB
D
Mn--
••
Proof, The system is observable if and only if the only vector that satisfies (3.45) is the zero vector. This is true if and only if ^ ( ^ ) = {0}, or if (3.47) is true. To determine the state XQ, apply (3.43) for ^ = 0 , 1 , . . . , n — 1, and rearrange in a form of a system of linear equations to obtain (3.48). • The matrix M^ defined above has the special structure of a ToepUtz matrix. Note that a matrix T is Toeplitz if its (/, 7)th entry depends on the value i — j ' , that is, T is "constant along the diagonals." Similarly to the continuous-time case, we now define the observability Gramian. DEFINITION 3.14. The observability Gramian of the system (3.41) is iho nxn matrix ^-1
(3.49) i=o
If ^k = [C^, ( C A ) ^ , . . . , (CA^-i)^]^ (with ^n = ^ ) , then Wo{0,K) = ^^^K-
(3.50)
The following result is apparent. LEMMA 3.13. ^ ( ^ ) = ^(Wo(0,K))
for every K>n.
Proof, The proof is left as an exercise for the reader.
•
COROLLARY 3.14. The system (3.41) is (completely state) observable, or the pair (A, C) is observable, if and only if rankWo{0,K)=n
(3.51)
for some (and consequently for all) K > n. If the system is observable, the state XQ at ^ = 0 is given by xo = W-\0,K)^^[Yo,K-i
-MkUo,K-i].
(3.52)
where K >n. Proof Statement (3.51) is a direct consequence of Corollary 3.12 and Lemma 3.13. To obtain (3.52), rewrite (3.48) in terms of K, premultiply by ^ ^ , and use relation (3.50).
259 CHAPTER 3 :
Controllability, Observability, and Special Forms
260
EXAMPLE 3.6. Consider the system in Example 1.4, x{k + 1) = Ax{k),y{k)
Linear Systems
Cx(k), where A = ' foYK
= n = 2is
1
1
and C = [0,1]. The observability Gramian
given by (3.49), Wo(0,2) =
ro 11 ro^ ro r ro^ ni 10, IJ .1 1. — .1. [0,1] + i_ [1,1] .1 iJ [i.
Xi^o(A^yC^CA^
0 0
0 1
1
1
=
Wo{0,K) [0, 1] +
1
1
, which is + ll 1 1 2 ro 1 is of full rank as of full rank. Therefore, the system is observable. Note that 0 =
Ll 1 well (see Example 1.4). Notice also that rank Wo(0, K) = 2 for every K > 2 and that
"0 1 "1 1 = 1 2 1 1
T r
0 1 0^0. The unique vector x(0) 1 1 can be determined from y(0) and yil) using (3.52) to obtain [refer to (3.50)] W^(0, 2)
xi(0) = w ; Ho, 2)0^ 3^(0) ^2(0)J 3^(1)
3^(1) - 3^(0)
This is the same as the result obtained in Example 1.4, using an alternative approach. Constructibility refers to the ability to determine uniquely the state x{0) from knowledge of current and past outputs and inputs. This is in contrast to observability, which utilizes future outputs and inputs. The easiest way to define constructibility is by the use of (3.44), where x(0) = XQ is to be determined from past data y{k), A: < 0. Note, however, that for fc < 0, A^ may not exist; in fact, it exists only when A is nonsingular. To avoid making restrictive assumptions, we shall define unconstructible states in a slightly different way than anticipated. Unfortunately, this definition is not very transparent. It turns out that by using this definition, an unconstructible state can be related to an unobservable state in a manner analogous to the way a controllable state was related to a reachable state in Section 3.2 (see also the discussion of duality in Section 3.1). DEFINITION 3.15. A State X is unconstructible if for every ^ > 0 there exists x i /?« such that =
Ak A'x,
Cx = 0.
(3.53)
The set of all unconstructible states, R^, is called the unconstructible subspace. The system (3.41) is (completely state) constructible, or the pair (A, C) is constructible, if the only state x E /?" that is unconstructible is x = 0, i.e., if R^ = {0}. • Note that if A is nonsingular, then (3.53) simply states that x is unconstructible if CA~^x = 0 for every k> 0 (compare this with Definition 3.13 of an unobservable state). The results that can be derived for constructibility are simply dual to the results on controllability. They are presented briefly below, but first, a technical result must be established. LEMMA 3.15. If X E M'(€) then Ax J{(€) is an A-invariant subspace.
>r(0), i.e., the unobservable subspace Ro =
Proof, Let x E >r(0), so that Ox = 0. Then CA^x = O f o r O < / : < n - l . This statement is also true for k> n — 1, in view of the Cayley-Hamilton Theorem. Therefore, OAx = 0, i.e.. Ax E M{p). •
THEOREM 3.16. Consider the systemx(^+1): --Ax{k)+Bu{k),y{k)=Cx{k)+Du{k) given in (3.41). (i) If a state x is unconstructible, then it is unobservable. (ii) RcnCRo. (iii) If the system is (completely state) observable, or the pair (A, C) is observable, then the system is also (completely state) constructible, or the pair {A,C) is constructible. If A is nonsingular, then relations (i) and (iii) are if and only if statements. In this case, constructibility also implies observability. Furthermore, in this case, (ii) becomes an equality, i.e., Rcw = RoProof, This theorem is dual to Theorem 2.22, which relates reachability and controllability in the discrete-time case. To verify (i), assume that x satisfies (3.53) and premultiply by C to obtain Cx = CA^x for every ^ > 0. Note that Cx = 0 since for k = 0^ x = x, and Cx = 0. Therefore, CA^x = 0 for every ^ > 0, i.e., x G ^ ( ^ ) . In view of Lemma 3.15, x = A^x G ./K(^), and thus, x is unobservable. Since x is arbitrary, we have also verified (ii). When the system is observable, Ro = {0}, which in view of (ii), implies that Ren = {0} or that the system is constructible. This proves (iii). Alternatively, one could also prove this directly: assume that the system is observable but not constructible. Then there exist x,x 7^ 0, which satisfy (3.53). As above, this implies that x G ./K(^), which is a contradiction since the system is observable. Consider now the case when A is nonsingular and let x be unobservable. Then, in view of Lemma 3.15, x = A~^x is also in ^ ( ^ ) , i.e., Cx = 0. Therefore, x = A^x is unconstructible, in view of Definition 3.15. This implies also that R^ C Ren, and therefore, Ro = RCE, which proves that in the present case constructibility also implies observability. • EXAMPLE 3.7. Consider the system in Example 1.5, x{k+l)= where A
Ax{k),y{k) = Cx{k),
and C = [1,0]. As shown in Example 1.5, rank^ = rank
1 <2 = n, i.e., the system is not observable. All unobservable states are of the form [Ol is a basis for yK{^) = Ro, the unobservable subwhere a ^ R since a
m
space. The observability Gramian foYK = n = 2is Wo(0,2) = C^C + (CA)^(CA) 1 0 1 0 2 0 . Note that a basis for yK(Wo(0,2)) is and 0 0 + 0 0 0 0 ^ ( ^ ) = J^{Wo{0,2)). This verifies Lemma 3.13. In Example 1.6 it was shown that all the states x that satisfy CA~^x = 0 for every
{[:]}
^ > 0, i.e., all the unconstructible states, are given by a
,a e R. This verifies
Theorem 3.16 (i) and (ii) for the case when A is nonsingular. EXAMPLE 3.8. Consider the system x(k+ 1) = Ax(k),y(k) = Cx{k), where A 0 0 1 0 and C = [1,0]. The observability matrix is of rank 1, and there1 0 0 0 fore, the system is not observable. In fact, all states of the form a states since <
> is a basis for
I are unobservable
J^{^).
To check the defining relations (3.53) must be used since A is 5ck constructibility, consti singular. Cx = [l,0]x = 0 implies x
Substituting into x = A^x, we obtain
261 CHAPTER 3 :
Controllability, Observability, and Special Forms
262 Linear Systems
for k = 0,x = X, and x = 0 for fc > 1. Therefore, the only unconstructible state is X = 0, which imphes that the system is constructible (although it is unobservable). This means that the initial state x(0) can be uniquely determined from past measurements. In fact, from x{k + 1) = Ax{k) and y{k) = Cx{k), we obtain x(0) xi(-l) X2(-l). Therefore, x(0)
0 xi(-iy
XI ( - 1 ) and y ( - l ) = C x ( - l ) = [1,0] X2(-l).
rxi(o)i _ X2(0)J =xi(-l).
0 y(-l)
DEFINITION 3.16. The constructibility Gramian is defined as (3.54)
We note that Wc^(0,^) is well defined only when A is nonsingular. The observability and constructibility Gramians are related by Wo(0,K) =
(3.55)
{A^)^Wcn{0,K)A^,
as can easily be verified. When A is nonsingular, the state XQ at ^ = 0 can be determined from past outputs and inputs in the following manner. We consider (3.44) and note that in this case y{k) = CA^jco is valid for ^ < 0 as well. This implies that 'CA-""' -l,-n
"Xo
(3.56)
Xo
CA-' with Y-i^-n = [j^(—n),.. . , j ^ ( — 1 ) ] ^ . Equation (3.56) must be solved for XQ. Clearly, in the case of constructibility (and under the assumption that A is nonsingular), the matrix ^A~^ is of interest instead of ^ [compare this with the dual results in (2.66)]. In particular, the system is constructible if and only if rank {^A~'^) = rank ^ = n. EXAMPLE 3.9. Consider the system in Examples 1.4 and 3.6, namely, x ( ^ + 1) = [0 ll and C = [0,1]. Since A is nonsingular, to Ax{k),y{k) = Cx{k), where A 1 1 CA-^ check constructibility we consider ^A , which has full rank. CA-' Therefore, the system is constructible (as expected), since it is observable. To determine x(0), in view of (3.56), we note that which
XI (0) X2(0),
namely, (A^)^Wc„(0,2)A
}'(-2)
ffA-^x{Q)
}'(-2)
}'(-2) 1 1
K-l)+K-2)
XI (0) X2(0)
, from
. It is also easy to verify (3.55), = W,(0,2).
•
263
PART 2 SPECIAL F O R M S FOR TIME-INVARIANT SYSTEMS
CHAPTERS:
3.4 SPECIAL FORMS
Controllability, Observability, and Special Forms
In this section, important special forms for the state-space description of timeinvariant systems are presented. These forms are obtained by means of similarity transformations and are designed to reveal those features of a system that are related to the properties of controllability and observability. In Subsection A, special state-space forms that display the controllable (observable) part of a system and that separate this part from the uncontrollable (unobservable) part are presented. These forms, referred to as the standard forms for uncontrollable and unobservable systems, are very useful in establishing a number of results. In particular, these forms are used in Subsection B to derive alternative tests for controllability and observability, and in Subsection C to relate state-space and input-output descriptions. (Additionally, these forms are further used in Chapters 4 and 5.) In Subsection D, the controller and observer state-space forms are introduced. These are useful in the study of state feedback and state estimators (to be addressed in Chapter 4), and in state-space realizations (to be addressed in Chapter 5).
A. Standard Forms for Uncontrollable and Unobservable Systems We consider time-invariant systems described by equations of the form X = Ax + Bu,
y = Cx + Du,
(4.1)
where A £ R""^", B G T?"^™, C £ RP^", and D £ RP^"". It was shown earlier in this chapter that this system is state reachable or controllable-from-the-origin if and only if the n X mn controllability matrix % =
[B,AB,...,A''-^B]
(4.2)
has full row rank n, i.e., rank % = n. Recall that 9l(^) = Rr is the reachable subspace, which contains all the state vectors that can be reached from the zero vector in finite time by applying an appropriate input. If the system is reachable (or controllable-from-the-origin), then it is also controllable (or controllable-to-theorigin), and vice versa (see Section 3.2). It was also shown earlier in this chapter that system (4.1) is state observable if and only if the pnX n observability matrix C CA
(4.3)
CA n-\ has full column rank, i.e., rank € = n. Recall that J{(€) = RQ is the unobservable subspace that contains all the states that cannot be determined uniquely from input
264 Linear Systems
and output measurements in finite time. If the system is observable, then it is also constructible, and vice versa (see Section 3.3). Similar results were also derived for discrete-time time-invariant systems described by equations of the form x(k +1) = Ax(k) + Bu(kl
y(k) = Cx(k) + Du(k)
(4.4)
in Sections 3.1, 3.2, and 3.3. Again, rank % = n and rank 0 = n are the necessary and sufficient conditions for state reachability and observability, respectively. Reachability always implies controllability and observability always implies constructibility, as in the continuous-time case. However, in the discrete-time case, controllability does not necessarily imply reachability and constructibility does not imply observability, unless A is nonsingular. Next, we will address standard forms for unreachable and unobservable systems both for the continuous-time and the discrete-time time-invariant cases. These forms will be referred to as standard forms for uncontrollable systems, rather than unreachable systems, and standard forms for unobservable systems, respectively. This is to conform with the established terminology in the literature, where the term "controllable" is used instead of "reachable," perhaps because of emphasis on the continuous-time case. It should be noted, however, that by the term controllable in this section we mean controllable-from-the-origin, i.e., reachable. 1. Standard form for uncontrollable systems If the system (4.1) [or (4.4)] is not completely controllable (-from-the-origin), then it is possible to "separate" the controllable part of the system by means of an appropriate similarity transformation. This amounts to changing the basis of the state space (see Section 2.2) so that all the vectors in the controllable (-from-the-origin) or reachable subspace Rr have certain structure. In particular, let rank ^ = nr < n, i.e., the pair (A, B) is not controllable. This implies that the subspace Rr = 2/l(^) has dimension n^. Let {vi, V2,..., v„J be a basis for Rr. These n^ vectors can be, for example, any nr linearly independent columns of ^ . Define the nX n similarity transformation matrix Q= [Vl,V2,...,Vn,,Qn-n^
(4.5)
where the nX (n - nr) matrix Qn-nr contains n - nr linearly independent vectors chosen so that Q is nonsingular. There are many such choices. We are now in a position to prove the following result. LEMMA 4.1. For (A, B) uncontrollable, there is a nonsingular matrix Q such that A = Q-'AQ =
Ai 0
An A2
and
B = Q'^B =
(4.6)
where Ai E R'^rXnr^ ^i G R'^rX'^^ and the pair (Ai, Bi) is controllable. The pair (A, B) is in the standard form for uncontrollable systems. Proof We need to show that = QA. 0 A2. Since the subspace Rr is A-invariant (see Lemma 2.15 in this chapter), Av/ G Rr, which can be written as a linear combination of only the nr vectors in a basis of Rr. Thus, AQ
= A[Vi, . . ., Vn,, Qn-nr]
= l^h • • •, V„,,
Qn-nr]
Ai in A is an rir X rir matrix, and the (n — rir) X rir matrix below it in A is a zero matrix.265 Similarly, we also need to show that CHAPTERS: Controllability, B = [vi,...,v„,, Q„-„J -QB. Observability, 0 and Special But this is true for similar reasons: the columns of 5 are in the range of'^ or in Rr. Forms The n X nm controllability matrix '^ of (A, B) is ^ = [B,AB,...,A"-'B] which clearly has ranfcC = that
=
Ar'Bi 0
0
(4.7)
0
rank[Bi,AiBi,... A"{''^Bi,...,A"^-^Bi]
= n,.Note (4.8)
The range of ^ is the controllable (-from-the-origin) subspace of (A, B). It contains vectorsonlyoftheform[a^, 0]^, wherea G JR"^ Since dim S/l(^) = rank% = rir, every vector of the form [a^, 0]^ is a controllable (state) vector. In other words, the similarity transformation has changed the basis of /?" in such a manner that all controllable (-from-the-origin) vectors, expressed in terms of this new basis, have this very particular structure of zeros in the last n — rir entries. Given system (4.1) [or (4.4)], if a new state x(t) is taken to be x(t) = Q~^x(t), then k = Ax + Bu,
y = Cx + bu,
(4.9)
where A = Q~^AQ,B = Q~^B,C = CQ, and D = D constitutes an equivalent representation (see Chapter 2). For Q as in Lemma 4.1, we obtain An
Xi X2
0
Ai.
u,y = [Ci, C2]
+ Du,
(4.10)
where x = {x[, X2] with xi G R^' and where (Ai, ^ i ) is controllable. The matrix C = [Ci, C2] does not have any particular structure. This representation is called a standard form for the uncontrollable system. The state equation can now be written as (4.11)
X\ ^ A i ^ i + BiW + A12X2, X2 = A2X2,
which shows that the input u does not affect the trajectory component X2(t) at all, and therefore, X2{t) is determined only by the value of its initial vector. The input u certainly affects xi (t). Note also that the trajectory component xi (t) is also influenced by X2(t). In fact. xi(t)
= ^^i^Jci(O) +
e'^'^'~^^Biu(T)dT +
oMit-T.-^Ane^^'dr
X2(0).
(4.12)
The nr eigenvalues of Ai and the corresponding modes are called controllable eigenvalues and controllable modes of the pair (A, B) or of system (4.1) [or of (4.4)]. The n- nr eigenvalues of A2 and the corresponding modes are called the uncontrollable eigenvalues and uncontrollable modes, respectively. It is interesting to observe that in the zero-state response of the system (zero initial conditions) the uncontrollable modes are completely absent. In particular.
266 Linear Systems
in the solution x(t) = e^^x(0) + \Q e^^^ ^'^Bu{T)dr of x = Ax + Bu, given x(0), notice that eMt-r)^ ^ [Qe^(^-^^Q-'][QB] = Q
0
(show this), where Ai [from (4.6)] contains only the controllable eigenvalues. Therefore, the input u(t) cannot directly influence the uncontrollable modes. Note, however, that the uncontrollable modes do appear in the zero-input response e^^x(0). The same observations can be made for discrete-time systems (4.4), where the quantity A^B is of interest (show this). ro EXAMPLE 4.1. Given A = 1
-1 11 "1 0" -2 1 and 5 = 1 1 we wish to reduce sysLO 1 -ij .1 2. tern (4.1) to the standard form (4.6). Here % = [B, AB, A^B]
1
0 :
1
1 :
1
2 :
1 : 0 -1 0 0 : 0 : 0
-1
1
and rank % = nr'=2<3 = n. Thus, the subspace Ry = S?l(^) has dimension rir = 2, and a basis {vi, V2} can be found by taking two linearly independent columns of ^, say, the first two, to obtain [Vl,V2, Gl] =
1 0
0
1 1
0
1 2 1 The third column of Q was selected so that Q is nonsingular. Note that the first two columns of Q could have been the first and third columns of ^ instead, or any other two linearly independent vectors obtained as a linear combination of the columns in %. For the above choice for Q we have.
ij [0
0 -1 1 -1 4 -2
0
-1
B= Q-^B =
1 -1 1
1
:
0 1 -2
01 n 0
ij
Ai
:
An
.0
:
M_
0
: -2_
0
1
11 ri 0 01 0 1 1 0 2] [1 2 i_
1 :
0
1] ri 0 0" 1 1 1 0 -ij [1 2 1.
01 To - 1 0 1-2
1 0 A = Q-'AQ = - 1 1 1 -2
0]
1 1 = 1 2^
" 1
0 '
0
1
0
0J
[B11
= 0
where (Ai, Bi) is controllable [verify this and show that ^ = Q ^^, i.e., verify (4.8)].
The matrix A has three eigenvalues at 0 , - 1 , - 2 . It is clear from (A, B) that the eigenvalues 0 , - 1 are controllable (in Ai), while - 2 is an uncontrollable eigenvalue (in A2). • 2. Standard form for unobservable systems The standard form for an unobservable system can be derived in a similar way as the standard form of uncontrollable systems. If the system (4.1) [or (4.4)] is not completely state observable, then it is possible to "separate" the unobservable part of the system by means of a similarity transformation. This amounts to changing the basis of the state space so that all the vectors in the unobservable subspace Ro have a certain structure. As in the preceding discussion concerning systems or pairs (A, B) that are not completely controllable, we shall presently select a similarity transformation Q to reduce a pair (A, C), which is not completely observable, to a particular form. This can be accomplished in two ways. The simplest way is to invoke duality and work with the pair (AD = A^,Bo = C^), which is not controllable (refer to the discussion of dual systems in Section 3.1). If Lemma 4.1 is applied, then AD
= Q'DADQD
=
AD\
0
ADU AD2
BD
-
QD BD
-
BDI
0
where (Aoi, Boi) is controllable. Taking the dual again, we obtain the pair (A, C), which has the desired properties. In particular, ' A^
A = A'^ = QhAUQor' = QoMGhr' = 1 _
c = Bi = BUQlr' = c(Qir
A^
0 A^
^D12
^D2
(4.13)
[B:IvOl
where (A^p Bj^^) is completely observable by duality (see Theorem 1.1). 0 1 0] we wish to reduce -1 - 2 1 andC = 1 1 -ij system (4.1) to the standard form (4.13). To accomplish this, let AD = A^ and BD = C^ • Notice that the pair (Ao, Bo) is precisely the pair (A, B) of Example 4.1. •
EXAMPLE 4.2. Given A =
A pair (A, C) can of course also be reduced directly to the standard form for unobservable systems. This is accomplished in the following. Consider the system (4.1) [or (4.4)] and the observability matrix 0 in (4.3). Let rank € = rio < n, i.e., the pair (A, C) is not completely observable. This implies that the unobservable subspace Ro = J^(€) has dimension n - HQ. Let {vi,..., v„-„J be a basis for 7?^ and define an n X n similarity transformation matrix Q as Q =
[Qno,Vi,...,Vn-nol
(4.14)
where the n^irio matrix Qn^ contains UQ linearly independent vectors chosen so that Q is nonsingular. Clearly, there are many such choices. LEMMA 4.2. For (A, C) unobservable, there is a nonsingular matrix Q such that \Ai 0 C = CQ = [Ci, 0], (4.15) and A = Q-'AQ = U21 A2
267 CHAPTER 3 :
Controllability, Observability, and Special Forms
268 Linear Systems
where Ai G /?«^^««, Ci E T^^^^^ and the pair (Ai, d ) is observable. The pair (A, C) is in the standard form for unobservahle systems. Proof We need to show that 0 A2.
A21
= eA.
Since the unobservahle subspace Ro is A-invariant (see Lemma 3.15), Av/ E T^^, which can be written as a linear combination of only the n- rio vectors in a basis of Ro. Thus, A2 in A is an (n ~ rio) X (n - Ho) matrix, and the rio X (n - rio) matrix above it in A is a zero matrix. Similarly, we also need to show that CQ = C[Q,„vi,...,v,_„J = [Ci,0] = C. This is true since Cvi = 0.
•
The pnX n observability matrix 0 of (A, C) is C CA
Ci CiAi
CA n-\
C,A\-'
(4.16)
which clearly has Ci CiAi rank€ = rank CiA «o-i
= ^«
ciAr Note that (4.17)
= €Q,
The null space of © is the unobservable subspace of (A, C). It contains vectors only of the form [0, a'^f, where a G /^"""o. Since dim >f(d) = n- rank € = n - no, every vector of the form [0, a ^ ] ^ is an unobservable (state) vector. In other words, the similarity transformation has changed the basis of R'^ in such a manner that all unobservable vectors expressed in terms of this new basis have this very particular structure—zeros in the first no entries. For Q chosen as in Lemma 4.2 and (4.9), it assumes the form Ai A21
0 A2
Bi u,y = [Ci, 0] Bi
+ Du,
(4.18)
where x = [x[, x^Y with xi G 7?"^ and where {A\, C\) is observable. The matrix B = [BJ, B^]^ does not have any particular form. This representation is called a standard form for the unobservable system. The no eigenvalues of Ai and the corresponding modes are called observable eigenvalues and observable modes of the pair (A, C) or of the system (4.1) [or of (4.4)]. The n - no eigenvalues of A2 and the corresponding modes are called unobservable eigenvalues and unobservable modes, respectively.
Notice that the trajectory component x(t), which is observed via the output 3;, is not influenced at all by X2, the trajectory of which is determined primarily by the eigenvalues of A2. The unobservable modes of the system are completely absent from the output. In particular, given x = Ax -\- Bu, y = Cx with initial state x(0), we have y{t)
Jo
and Ce"^' = [CQ-^][Qe^'Q-^] = [Cie'^^',0]Q-^ (show this), where Ai [from (4.15)] contains only the observable eigenvalues. Therefore, the unobservable modes cannot be seen by observing the output. The same observations can be made for discrete-time systems where the quantity CA^ is of interest (show this). EXAMPLE 4.3. Given A = \ ^ 2
^ | and C = [1, 1], we wish to reduce system -3J
(4.1) to the standard form (4.15). To accomplish this, we compute 0 = , which has rank € = rio = 1 < 2 = n. Therefore, the unobservj space Ro = M(€) has dimension n - Uo = 1. In view of (4.14), Q = [Qhvi] =
0
:
1
1 : -ij
where vi = [1, - 1 ] ^ is a basis for Ro, and Qi was chosen so that Q is nonsingular. Then A = Q-'AQ
C= CQ= [1,1]
1 1 1 0
"0 1
1 -1
Ai
0
A21
A2,
[1,0] = [Ci,0],
where (Ai, Ci) is observable [show this and verify that 6 = €Q, i.e., verify (4.17)]. The matrix A has two eigenvalues at - 1 , - 2 . It is clear from (A, C) that the eigenvalue - 2 is observable (in Ai), while - 1 is an unobservable eigenvalue (in A2). • 3. Kalman's Decomposition Theorem Lemmas 4.1 and 4.2 can be combined to obtain an equivalent representation of (4.1) where the reachable and observable parts of this system can readily be identified. To this end, we consider system (4.9) again and proceed, in the following, to construct the nX n required similarity transformation matrix Q. As before, we let rir denote the dimension of the controllable (-from-the-origin) subspace 7?r i-e., ^r = dim Rr = dim9l(^) = ran^^. The dimension of the unobservable subspace/?^ = >r(0)isgivenby n^ = n-rankG = n-n^.Let/i^^bethe dimension of the subspace Rro = Rr(^ Ro that contains all the state vectors x E R^ that are controllable but unobservable. We choose Q = [Vl, . . ., Vn,-nro + h . • ., Vnr^ QN, Vi, . . ., V ^ , - « , J ,
(4.19)
269 CHAPTERS:
Controllability, Observability, and Special Forms
270 Linear Systems
where the rir vectors in { v i , . . . , v ^ ^ } form a basis for Rr. The last riro vectors {^nr-nro+ii"'i^nr} ii^ the basis for Rr dire chosen so that they form a basis for Rro = Rr^Ro' The Ho — Uro = {n — HQ — Uro) vcctors { v i , . . . , Vn^-Tiro) ^^^ Selected so that when taken together with the rird vectors {v^^_^^-+i,..., V^^} they form a basis for Ro, the unobservable subspace. The remaining N = n— {fir + rio — Uro) columns in Q^ are simply selected so that Q is nonsingular. The following theorem is called the Canonical Structure Theorem or Kalman's Decomposition Theorem. THEOREM 4.3. For (A,5) uncontrollable and {A,C) unobservable, there is a nonsingular matrix Q such that
A = Or^AQ =
All
0
A21 0 0
A22 0 0
Al3 A23 A33 A43
'B{
0 " A24 0 A44_
B = Q-'B =
B2 0 0
1
(4.20)
C = Ce=[Ci,0,C3,0], where (i)
(A„5,)with Ac^
(ii)
"All
0 "
A21
A22_
and
Be =
B{
is controllable (-from-the-origin), where A^ G R^'^^',Bc G R^^^^, {Ao,Co) with \An 0
Ai31 A33
and
Co = [Ci,C3]
is observable, where Ag G R^o><no ^^^ Q ^ j^pxno ^j^^^ where the dimensions of the matrices Aij.Bi, and Cj are A l l : {nr - Uro) X {rir - Uro),
A33 : {n-{nr + no-nro)) B\ : {fir — firo) X m, Ci : px(nr-nro), (iii)
A22 : Uro X Uro,
x {n - {ur + no - firo)), A44 : {fio-firo) x {fio-firo), B2 '. Uro X m, C3 : p x (n-(rir + no-nro)),
the triple (Ai 1, 5 i , Ci) is such that (Ai 1,5i) is controllable (from-the-origin) and (Aii,Ci) is observable.
Proof, For details of the proof, refer to [8] and to R. E. Kalman, "On the Computation of the Reachable/Observable Canonical Form," SI AM J. Control and Optimization, Vol. 20, No. 2, pp. 258-260, 1982, where further classifications to [8] and an updated method of selecting Q are given. • The similarity transformation (4.19) has altered the basis of the state space in such a manner that the vectors in the controllable (-from-the-origin) subspace Rr, the vectors in the unobservable subspace RQ, and the vectors in the subspace Rro^Ro all have specific forms. To see this, we construct the controllability matrix C = [ 5 , . . . ,A^~^5] whose range is the controllable (-from-the-origin or reachable) subspace and the observability matrix S = [ C ^ , . . . , (CA^~^)^]^, whose null space is the unobservable subspace. Then, all controllable states are of the form [^^,^2 , 0 , 0 ] ^ , all the unobservable ones have the structure [0,^2 , 0 , ^ 4 ] ^ , while states of the form [0,^2 , 0 , 0 ] ^ characterize Rro, i.e., they are controllable but unobservable.
Similarly to the previous two lemmas, the eigenvalues of A, or of A, are the eigenvalues of An, A22, A33, and A44, i.e., |A/ - A\ = |A/ - A\ = \XI - AiillA/ - A22IIA/ - A33IIA/ - A44I.
(4.21)
If in particular we consider the representation {A, B, C, D} given in (4.9), where Q was selected as in the canonical structure theorem given above, then
-1
0 1 pi
Ai3 An 0 A21 A22 A23 0 0 A33 [ 0 0 A43
r^ii
A24 U2 + B2 0 p3 0 A44J [x4 0
Xl
y = [Ci.0,C3,0]
X2
+ Dw.
(4.22)
This shows that the trajectory components corresponding to X3 and X4 are not affected by the input u. The modes associated with the eigenvalues of A33 and A44 determine the trajectory components for X3 and f 4 (compare this with the results in Lemma 4.1). Similarly to Lemma 4.2, the trajectory components for X2 and X4 cannot be observed from y, and are determined by the eigenvalues of A22 and A44. The following is now apparent (see also Fig. 3.4): The eigenvalues of All are controllable and observable, A22 are controllable and unobservable, A33 are uncontrollable and observable, A44 are uncontrollable and unobservable.
CO
I
FIGURE 3.4 Canonical decomposition (c and c denote controllable and uncontrollable, respectively). The connections of the c/c and 0/0 parts of the system to the input and output are emphasized. Note that the impulse response (transfer function) of the system, which is an input-output description only, represents the part of the system that is both controllable and observable (see Section 3.4C).
271 CHAPTERS:
Controllability, Observability, and Special Forms
272
Linear Systems
EXAMPLE 4.4. Given
ro - 1 1 "1 0" 1 -2 1 , 5 = 1 1 , and C = [0,1, 0], we wish to reduce system [o 1 -1 .1 2_ (4.1) to the canonical structure (or Kalman decomposition) form (4.20). The appropriate transformation matrix Q is given by (4.19). The matrix ^ was found in Example 4.1 and 0 1 0 C CA 1 -2 1 A-
4
2
-2
A basis for R^ = >r(0) is {(1, 0, -1)^}. Note that rir = Zrio = 1, and riro = 1- Therefore,
Q = [Vl, V2, QN]
1
1
0
1
0
0
1
-1
1
is an appropriate similarity matrix (check that det Q ¥- 0). We compute 0 1 01 ro 1 11 1 1 0 A = Q-'AQ 2 1 1 0 0 1 1 -1 0 1 -2 1 -ij 1 -1 ij [o 1
-1
" 1
0 ^ n-^n = 1 B = Q-'B 1
1 -1 -2
01 ri 0 1
0'
ij Ll
2.
1
All
0
Ai3
A21 0
A22 0
A23 A33
1 ' Bi'
=
0
-1
0
0
= B2 .0.
1 1 0 C = CQ = [0, 1, 0] 1 [1,0,0] = [Ci,0,C3]. 0 0 1 -1 1 The eigenvalue 0 (in An) is controllable and observable, the eigenvalue - 1 (in A22) is controllable and unobservable, and the eigenvalue - 2 (in A33) is uncontrollable and observable. There are no eigenvalues that are both uncontrollable and unobservable. • and
B. Eigenvalue/Eigenvector Tests for Controllability and Observability There are tests for controllability (-from-the-origin) and observability for both continuous- and discrete-time time invariant systems that involve the eigenvalues and eigenvectors of A. Some of these criteria are called PBH tests, after the initials of the codiscoverers (Popov-Belevitch-Hautus) of these tests. These tests are useful in theoretical analysis, and in addition, they are also attractive as computational tools.
THEOREM 4.4. (i) The pair (A, B) is uncontrollable if and only if there exists a 1 X ^ (in general) complex vector vt # 0 such that v / [ A / / - A , 5 ] = 0, where A/ is some complex scalar. (ii) The pair (A, C) is unobservable if and only if there exists annX complex vector v/ T^ 0 such that [A// - A] C
0,
(4.23) I (in general)
(4.24)
where A/ is some complex scalar. Proof, Only part (i) will be considered since (ii) can be proved using a similar argument, or directly, by duality arguments. (Sufficiency) Assume that (4.23) is satisfied. In view of v/A = A/V/ and ViB = 0, ViAB = XiViB = 0, and vtA^B = 0, ^ = 0, 1, 2 , . . . . Therefore, Vi% = Vi[B, AB,..., A^~^B] = 0, which shows that (A, B) is not completely controllable. (Necessity) Let (A, B) be uncontrollable and assume without loss of generality the standard form for A and B given in Lemma 4.1. We will show that there exist A/ and v/ so that (4.23) holds. Let Aj be an uncontrollable eigenvalue and let v/ = [0, a], a^ G C"~"% where a(XiI - A2) = 0, i.e., a is a left eigenvector of A2 corresponding to A/. Then v/[A,/ -A,B] = [0, a(\il - A2), 0] = 0, i.e., (4.23) is satisfied. • COROLLARY 4.5. (i) The pair (A, B) is controllable if and only if no left eigenvector of A is orthogonal to all the columns of B. (ii) The pair (A, C) is observable if and only if no right eigenvector of A is orthogonal to all the rows of C. Proof, The proof follows directly from Theorem 4.4.
•
COROLLARY 4.6. (i) A/ is an uncontrollable eigenvalue of (A, B) if and only if there exists a 1 X /I (in general) complex vector v/ 7^ 0 that satisfies (4.23). (ii) A/ is an unobservable eigenvalue of (A, C) if and only if there exists an /i X 1 (in general) complex vector v, 7^ 0 that satisfies (4.24). Proof Only part (i) will be considered, since part (ii) can be proved using a similar argument or directly, by duality arguments. (Sufficiency) Assume that (4.23) is satisfied. Now v/[A// - A, B] = 0, in view of the sufficiency proof of Theorem 4.4, implies that v/^ = v/[5, AB,..., A^'^B] = 0. Therefore, (A, B) is not controllable. Without loss of generality, assume that (A, B) is in the standard form of Lemma 4.1. In this case the controllability matrix has the form (4.7) with its top nr rows linearly independent and its lower n - nr rows being zero. Therefore, in view of Vj^ = 0, v/ has the form v/ = [0, a\ for some a E C"~"^ Now v/(A// - A) = 0 implies that a (A// - A2) = 0, which shows that A/ is an eigenvalue of A2, i.e., it is an uncontrollable eigenvalue. (Necessity) Let A/ be an uncontrollable eigenvalue of (A, B). Assume without loss of generality that the pair (A, B) is in the standard form of Lemma 4.1. Then v/ = [0, a ], where a i s s u c h t h a t a ( A / / - A2) = 0 (see the proof of Theorem 4.4), satisfies v/(A//- A) = 0. Also, ViB = [0, a]
0
= 0. So (4.23) is satisfied.
0 -1 1 "1 0' 1 -2 1 ,B = 1 1 , and C = [0,1, 0], as in 0 1 -1 .1 2. Example 4.4. The matrix A has three eigenvalues, Ai = 0, A2 ^ - 1 , and A3 = - 2 , with
EXAMPLE 4.5. Given are A =
273 CHAPTER 3 :
Controllability, Observability, and Special Forms
274 Linear Systems
corresponding right eigenvectors vi = [1,1,1]^, V2 = [1,0, - 1 ] ^ , V3 = [1,1, - 1 ] ^ and with left eigenvectors vi = [^,0, ^],V2 = [1, ~ 1 , 0], and V3 = [-^, 1, - 5 ] , respectively. In view of Corollary 4.6, viB = [1,1] 7^ 0 implies that Ai = 0 is controllable. This is because vi is the only nonzero vector (within a multiplication by a nonzero scalar) that satisfies vi(Ai/ - A) = 0, and so Vi5 7^ 0 impUes that the only 1 x 3 vector a that satisfies a[\il - A, B] = 0 is the zero vector, which in turn imphes that Ai is controllable in view of (i) of Lemma 4.6. For similar reasons Cvi = 1 7^ 0 imphes that Ai = 0 is observable; see (ii) of Lemma 4.6. Similarly, V2B = [0, - 1 ] 7^ 0 implies that A2 = - 1 is controllable, and Cv2 = 0 implies that A2 = - 1 is unobservable. Also, v^^B = [0, 0] implies that A3 = - 2 is uncontrollable, and CV3 = 1 7^ 0 implies that A3 = - 2 is observable. These results agree with the results derived in Example 4.4. • COROLLARY 4.7. (RANK TESTS), (ia) The pair (A, B) is controllable if and only if (4.25)
rank [A/ - A, 5] = n for all complex numbers A, or for all n eigenvalues A/ of A. (ib) Xi is an uncontrollable eigenvalue of A if and only if rank [A// - A, 5] < n.
(4.26)
(iia) The pair (A, C) is observable if and only if rank
XI - A C
(4.27)
for all complex numbers A, or for all n eigenvalues A/. (iib) Xi is an unobservable eigenvalue of A if and only if rank
\XiI - A] < n. C
(4.28)
Proof. The proofs follow in a straightforward manner from Theorem 4.4. Notice that the only values of A that can possibly reduce the rank of [XI - A, B] are the eigenvalues ofA. • EXAMPLE 4.6. If in Example 4.5 the eigenvalues Ai, A2, A3 of A are known, but the corresponding eigenvectors are not, consider the system matrix
P(s) =
si - A -C
B 0
s
1
-1
1
0
-1
s+2
-1
1
1
0
-1
s+\
1
2
-1
0
0
0
and determine rank [XJ - A, B] and rank -1 -1 = rank 0 0
1 1 -1 1
-2 rank [si — A, B]s=x^ = rank - 1 0
2 0 -1
rank
and
Xil - A . Notice that C
si - A C
5 = ^2
= 2<3
-1 -1 -1
= n
= 2<3
= n.
In view of Corollary 4.7, A2 = - 1 is unobservable and A3 = - 2 is uncontrollable. Verify that these are the only uncontrollable and unobservable eigenvalues by applying the rank tests of Corollary 4.7. Compare these results with the results in Example 4.5. E X A M P L E 4.7.
Let
and B
A
1 the eigenvalues
with A
of A. We would like to determine which of the eigenvalues are uncontrollable. Note that (A, B) is in the standard form for uncontrollable systems of Lemma 4.1, namely. Ai An . We know by inspection that the eigenvalue A2 = 1 of A2, is uncon0 Ai, Uz-l -1 11 trollable. Presently, [A// - A, B] and for V2 = [0, 1], V2[A2/ [ 0 A, - 1 oj A, B] = [0, 0], which in view of Theorem 4.4 and Corollary 4.6, imphes that A2 = 1 is an uncontrollable eigenvalue and that V2 = [0, 1] is the corresponding left eigenvector. Note that V2 is of the form [0, a], as discussed in the proof of Theorem 4.4. The other eigenvalue, Ai = 1, is controllable. It is the eigenvalue of Ai = 1, where (Ai, Bi) is controllable. Note that the corresponding eigenvector to the controllable eigenvalue is vi = [0,1], the same as V2- Therefore, when using the eigenvalue/eigenvector tests, one can detect in the present example only that at least one of the multiple eigenvalues is uncontrollable; this test is unable to detect that the other eigenvalue is controllable. This situation arises when there are multiple eigenvalues, in which case care should be taken when using the eigenvalue/eigenvector tests. When the eigenvalues are distinct, then each of them can specifically be identified as being controllable or uncontrollable by the eigenvalue/eigenvector test. For another example, try A =
1
0"
0
1.
and 5 =
1' 0.
C. Relating State-Space and Input-Output Descriptions The system x = Ax-^ Bu, y = Cx-^ Du given in (4.1) has pX m transfer function matrix H(s) = C(sl - A)-^B + D = C(sl - Ay^B + A
(4.29)
where {A, B, C, D} is the equivalent representation given in (4.9) (see also Sections 2.5 and 2.6). Consider now the Kalman Decomposition Theorem and the representation (4.22). We wish to investigate which of the submatrices A/y, Bi, Cj determine H(s) and which do not. The inverse of si - A can be determined by repeated application of the formulas -1
0
8
la 0" [y 8
a §-1
0 -1
a
0 (4.30) -1 ya where a, (5, 7, 8 are matrices, with a and 8 square and nonsingular. It turns out (verify) that and
H{s) = Ci(sl - Aii)"iJ5i + A
(4.31)
that is, the only part of the system that determines the external description is {All, Bi, Ci, D}, the subsystem that is both controllable and observable [see Theorem 4.3(iii)]. Analogous results exist in the time domain. Specifically, taking the
275 CHAPTERS:
Controllability, Observability, and Special Forms
276 Linear Systems
inverse Laplace transform of both sides in (4.29), the impulse response of the system for r > 0 is derived as (see Chapter 2) H{t, 0) = de^'^'Bi
+ D8{t),
(4.32)
which depends only on the controllable and observable parts of the system, as expected. Similar results exist for discrete-time systems described by (4.4). For such systems, the transfer function matrix H{z) and the pulse response H{k, 0) (see Chapter 2) are given by
and
(4.33)
Cx{zI-Axi)-^Bi+D
H{z) H{k, 0)
k>0, k = 0,
D,
(4.34)
These depend only on the part of the system that is both controllable and observable, as in the continuous-time case. EXAMPLE 4.8. For the system x = Ax + Bu, y = Cx, where A, B, C are as in Examples 4.4 and 4.5, we have i/(5) = C{sI-A)-^B = Ci(sI-An)~^Bi = (1)(1/^)[1,1] = [l/s, l/s]. Notice that only the controllable and observable eigenvalue of A, Ai = 0 (in All), appears in the transfer function as a pole. All other eigenvalues (A2 = - 1 , A3 = - 2 ) cancel out. • EXAMPLE 4.9. The circuit depicted in Fig. 3.5 is described by the state-space equations 1 (RiC)
0
i(t) =
Ri
1 (RiC)
X2(t)
0 1
Xi(t)
L 1
v(0
1
Xi(t)
MO
where the voltage v(t) and current i(t) are the input and output variables of the system, xi(t) is the voltage across the capacitor, and X2(t) is the current through the inductor. We have i(s) = H(s)v(s) with the transfer function given by His) = C(sl - AT^B + D = (^'.C-L)s^(R,-R2) ^^ ^ ^ (Ls + R2){R\Cs + Ri)
^ 1 Ri
The eigenvalues of A are Ai = -l/(RiC) and A2 -R2IL. Note that in general [A// - A rank [A// —A,B] = rank = 2 = n, i.e., the system is controllable and
C
^2(0
/•(O
^i(0 v{t)
FIGURE 3.5
observable, unless the relation R1R2C = Lis satisfied. In this case Ai and the system matrix P(s) assumes the form
'^T P(s) =
si - A -C
B = D
0
-k
-Ri/L
Controllability, Observability, and Special Forms
1
^
1 1
-
In the following, assume that R1R2C = Lis satisfied. (i) Let Ri 7^ R2 and take R2 L^h V2I — .1
Ri 1 .'
- rvi
IM1~^
—
i<2 - ~Ri
1 1
-R, R2.
to be the linearly independent right and left eigenvectors corresponding to the eigenvalues Ai = A2 = -R2IL. The eigenvectors could have been any two linearly independent vectors since XJ - A = 0. They were chosen as above because they also have the property that V2^ = 0 and Cv2 = 0, which in view of Theorem 4.4, implies that A2 = -R2IL is both uncontrollable and unobservable. The eigenvalue Ai = -R2IL is rD
D 1
both controllable and observable since it can be seen using 2 = K ^\x.o reduce the representation to the canonical structure form (Kalman Decomposition Theorem) (verify this). The transfer function is in this case given by His)
{s + RxlL){s + R2IL) Ri(s + R2/L)(s + R2/L)
s + RilL Ri(s + R2/Ly
that is, only the controllable and observable eigenvalue appears as a pole in H(s), as expected. (ii) Let Ri = R2 = R and take Lvi, V2J
[Vl,V2] ^ =
In this case viB = 0 and Cv2 = 0. Thus, one of the eigenvalues, Ai = -RIL, is uncontrollable (but can be shown to be observable) and the other eigenvalue, A2 = -RIL, is unobservable (but can be shown to be controllable). In the present case, none of the eigenvalues appear in the transfer function. In fact.
His) = 1. as can readily be verified. Thus, in this case the network behaves as a constant resistance network. At this point it should be made clear that the modes that are uncontrollable and/or unobservable from certain inputs and outputs do not actually disappear; they are simply invisible from certain vantage points under certain conditions. (The voltages and currents of this network in the case of constant resistance [H{s) = l/R] are studied in Exercise 3.26.) • EXAMPLE 4.10. Consider the system i = Ax+Bu,y
B =
277 CHAPTERS:
RiL
^ .+
A2
= Cx where A
1 0 0 0-2 0 0 0-1
and C = [1, 1,0]. Using the eigenvalue/eigenvector test it can be shown
278
Linear Systems
(verify this) that the three eigenvalues of A (resp., the three modes of A) are Ai = 1 (resp., e^), which is controllable and observable, A2 = - 2 (resp., e'^^), which is uncontrollable and observable, and A3 = - 1 (resp., e^^), which is controllable and unobservable. The response due to the initial condition x(0) and the input u(t) is x(t) = e'''xiO)+ \ e''^'-^^Bu(T)dT 0" 0 x(0) + ,-t
and
KO = C^^'x(O) +
ft
- e«-T) -
U{T) dr 0 Jo _e-('--\
Ce^^'~^^Bu(T)dr
= [e\e~^\Q]x{0)+\
e^'''h(T)dT.
Notice that only controllable modes appear in e^^B [resp., only controllable eigenvalues appear in (si - A)~^5], only observable modes appear in Ce^^ [resp., only observable eigenvalues appear in C(5/ - A) ~ ^ ], and only modes that are both controllable and observable appear in Ce^^B [resp., only eigenvalues which are both controllable and observable appear in C{sl - A)~^B = H{s)]. For the discrete-time case, refer to Exercise 3.17d. •
D. Controller and Observer Forms It has been seen several times in this book that equivalent representations of systems = Ax + Bu,
y = Cx + Du,
(4.35)
X = Ax + Bu,
y = Cx + Du,
(4.36)
given by the equations
where x = Px,A = PAP-\ B = PB,C = CP-\ and D = D, may offer advantages over the original representation when P (or Q = P~^) is chosen in an appropriate manner. This is the case when P (or Q) is such that the new basis of the state space (see Section 2.2) provides a natural setting for the properties of interest. As a specific case, refer for example to Subsection 3.4A, where Q (and the new basis) was chosen so that the controllable and uncontrollable parts of the system were separated. The same results of course apply to discrete-time systems (4.4). This subsection shows how to select Q when (A, B) is controllable [or (A, C) is observable] to obtain the controller and observer forms. These special forms are very useful, especially when studying state-feedback control (and state observers) discussed in Chapter 4 and in realizations discussed in Chapter 5. These special forms are also very useful in establishing a quick way to shift between state-space representations and another very useful class of equivalent internal representations, the polynomial matrix representations studied in Chapter 7. Controller forms are considered first. Observer forms can of course be obtained directly in a similar manner as the controller forms, or they may be obtained by duality. This is addressed in the latter part of this section.
1. Controller forms
279
The controller form is a particular system representation where both matrices (A, B) have certain special structure. Since in this case A is in the companion form (see Section 2.2 in Chapter 2) the controller form is sometimes also referred to as the controllable companion form. Consider the system X = Ax-\- Bu,
y = Cx-\- Du,
where A G /?^x^ B G R'''''^, C G 7?^>'^ and D G lable (-from-the-origin). Then ran ^ ^ = n, where % =
RP"""^
[B,AB,...,A''~^Bl
(4.37)
and let (A, B) be control(4.38)
Assume that rank B = m ^ n.
(4.39)
Under these assumptions, r a n ^ ^ = n and rank B = m. We will show how to obtain an equivalent pair (A, B) in controller form, first for the single-input case (m = 1) and then for the multi-input case (m > 1). Before this is accomplished, we discuss two special cases that do not satisfy the above assumptions that rank B = m and that (A, B) is controllable. 1. If the m columns of 5 are not linearly independent (rankB = r < m), then there exists an m X m nonsingular matrix K (or equivently, there exist elementary column operations) so that BK = [Br, 0], where the r columns of Br are linearly independent (ran ^ 5^ = r). Note that i = Ax + Bu = Ax + {BK){K~^u) = Ax^[Br, 0]
Ur Uyyi—y
= Ax + BrUr, which clearly shows that when rankB = r < m the
same input action to the system can be accomplished by only r inputs, instead of m inputs, and there is a redundancy of inputs, which in control problems clearly implies that a reconsideration of the input choices is in order. The pair (A, Br), which is controllable when (A, B) is controllable (show this), can now be reduced to controller form, using the method developed below. 2. When (A, B) is not completely controllable, then a two-step approach can be taken. First, the controllable part is isolated (see Subsection 3.4A) and then is reduced to the controller form, using the methods of this section. In particular, consider the system x = Ax + Bu with A G R"^^"", B G i^^X'", and rankB = m. Let rank [B, AB,..., A^~^B] = nr < n. Then there exists a transformation Pi such Ai Al2 that PiAPand PiB = where Ai G P^^^^^^ Bi G P'^^^'^, 0 A2 and (Ai,Bi) is controllable (Subsection 3.4A). Since (Ai, Bi) is controllable, there exists a transformation P2 such that P2A1P2 ^ = Mc, and P2B1 = Bic, where Aic, Bic is in controller form, defined below. Combining, we obtain PAp-'- 1 =^ and
Ale
0
"B = Bic 0
PiAn A2 _
(4.40)
CHAPTER 3 :
Controllability, Observability, and Special Forms
280
Linear Systems
[where Axc ^ R'^'^'^'.Bic G R'^rxm^ ^^^ (Aic,Bic) is controllable], which is in controller form. Note that P2 0
0 / Pi-
(4.41)
Single-input case (m = 1). The representation {Ac^Bc^Cc^Dc} in controller A = PAP-^ and Bc= B = PB with form is given by Ac
0 -oco
(4.42)
Br
0 -ai
1 -OCn-l
where the coefficients ai dire the coefficients of the characteristic polynomial a{s) of A, that is, a{s)= det {si - A) = s"" ^ an-is""'^ ^ h^i^ + ao(4.43) Note that Q = C = CP~^ and Dc = D do not have any particular structure. The structure of (Ac,Be) is very useful (in control problems) and the representation {Ac^Bc^Cc^Dc} shall be referred to as the controller form of the system. The similarity transformation matrix P is obtained as follows. The controllability matrix ^ = [B^AB,... ,A^ ^5] is in this case minxn where q is the nth row of ^
nonsingular matrix and ^
and x indicates the remaining entries of "^
. Then
qA (4.44)
qAn-\ To show that PAP~^ = Ac and PB = Be given in (4.42), note first that qA'~^B = 0, / = 1,... , n - 1, and qA^'-^B = 1. This can be verified from the definition of q, which implies that q^ = [0,0,..., 1] (verify this). Now "0 0 P ^ = P[5,A5,...,A^"^5]
0 0
•• ••
1
r
X
•
^ c -
(4.45)
1 1
X
X
X
which imphes that | P ^ | = | P | | ^ | 7^ 0 or that |P| ^ 0. Therefore, P quaUfies as a similarity transformation matrix. In view of (4.45), PB = [0,0,...,!]^ = Be. Furthermore, qA AcP
qAn-l qA^
PA.
(4.46)
where in the last row of A^P, the relation - X/^=o ^/^^ ^ ^ " was used [which is the Cay ley-Hamilton Theorem, namely, a(A) = 0]. EXAMPLE4.il. Let A =
0 1 0
1 0 0
l^-/ - A| = (^ + 1)(5' - l)(s + 2) 0 by A, = 0 2
= s^ +2s^ - s -2, {Ac, Be} in controller form is given 1 0 and Br = 0 1 1 -2
1 - I -1 _
2
2
The third (the nth) row of^ ^ is q = [- ^, - ^, ^], and therefore, i
P^
3 2 3
1
qA qA^
2
i L
4 3J
2
It can now easily be verified that Ac = PAP ^ or = PA,
AcP
and that Be = P5. [Verify that P^ = %c is given by (4.45) and also compare with Exercise 3.23, which explicitly derives ^cJ • An alternative form to (4.42) is -ax 0
-Oin-\
1 Aci =
-ao 0 ^cl
0
•••
1
-
(4.47)
0
which is obtained if the similarity transformation matrix is taken to be r^A^-i" A
P^ =
CHAPTER 3 :
Controllability, Observability, and Special Forms
" 1" 0" 0 and B = - 1 Since n = 3 and -2. 1.
The transformadon matrix P that reduces (A, B) to (Ac = PAP ^,Bc = PB) is now derived. We have , 1 - 1 n = [B, AB, A^B] = -1 - 1 1 -2 and
281
(4.48) qA
i.e., by reversing the order of the rows of P in (4.44). The reader should verify this (see Exercise 3.25).
282 Linear Systems
0 / 'x x" and Br X X or / 0 has the form [0 0 . . . 1]^ or [1 0 . . . 0]^, respectively. These representations are very useful when studying pole assignment via state feedback control law and are used in the next chapter. In the above, Ac is a companion matrix of the form
A companion matrix could also be of the form
"0 x' or 'x
0"
(see Section / X X / 2.2) with the coefficients -[ao,..., c^n-iV in the last or the first column. It is shown here, for completeness, how to determine controller forms where Ac are such companion matrices. In particular, if G2 = Pi^ =
then
Ac2 =
Qi'^Qi =
"0 ••• 0 1 ••• 0 0
•••
(4.49)
[B,AB,... ,A"-^5] = -ao Bc2 = Q2'B =
'1 0
(4.50)
0
1
Also, if Q3 = P3' =
then
Ac3 -
-a„-i
1
...
0"
-ao
0 0
... ...
1 0
Q^'AQ^ =
(4.51)
• .,Bl
[A"~'B,
"0 Bc3 = Q^'B =
0 1
(4.52)
(Ac, Be) in (4.50) and (4.52) are also in controller canonical or controllable companion form. (The reader is encouraged to verify these expressions. See also Exercise 3.25.) We also note that if the structures of Ac and Be are specified, then P is uniquely determined (see Exercise 3.24). That is, given (A^ Be) in any of the above four controllable companion forms, P is readily uniquely determined in each case by P = ^ ^ " ^ assuming that P also satisfies PA^B ^ A^Bc (in view of Exercise 3.24), which it does in the above four cases. If different Be are desirable, then an appropriate P can be found by the same formula. Note that ^ denotes the controllability matrix of (Ac, Be). r-1 0 0 , as in Example 4.11. AI0 1 0 and 5 = 0 0 -2 temative controller forms can be derived for different P. In particular, if
EXAMPLE 4.12. Let A =
\qA^ (i)P = P, = \ qA Lq J in Example 4.11), then
1 2 1 2 1 2
1 6 1 6 1 6
- 2 1 2" 1 0 0 , 0 1 0_
4-1 3 2 3 1 3-1
as in (4.48) {%^ ', and q were found
Be
-
1 2 1
as in (4.47). Note that in the present case AciPi = Bel =
1 6 1 6 1 6
]
- 2
PiB.
" 1 - 1 (ii) 22 = ^ = - 1 - 1 . 1 - 2 Ac2 =
"0 1 .0
8 3 4 3 2 3
283 PiA,
r - 1 , as in (4.49). Then 4. 2" 0 0 1 > 1 --2.
ri] Bc2 = Q2'B
=
0 LOJ
as in (4.50). " 1 (iii)e3 = [A^B,AB,B] = - 1 4 "-2 1 Ac3 = 2 -1 as in (4.52). Note that gsA^s = - 1 .-8
-1 -1 -2 1 0 0
1] - 1 , as in (4.51). Then 1. 0" roi 1 Bc3 - 0 , 0. [ij
1 -1" -1 -1 4 - 2_
= AGs, Q3Bc3 =
r 1" -1 = L 1.
Multi-input case (m > I). In this case the n X mn matrix % given in (4.38) is not square, and there are typically many sets of n columns of % that are linearly independent (rank% = n). Depending on which columns are chosen and in what order, different controller forms (controllable companion forms) are derived. Note that in the case when m = 1, four different controller forms were derived, even though there was only one set of n linearly independent columns. In the present case there are many more such choices. The form that will be used most often in the following is a generalization of (A^ Be) given in (4.42), and this is the form that will be derived first. Other forms will be discussed as well. Let A = PAP~^ and B = PB, where P is constructed as follows: consider
= [bu ,,.,bm,
Ah,, . . ., Abn,, . . ., A^-'b,,
. . .,
A^-'bml
(4.53)
where the bi,.. .,bm are the m columns of 5. Select, starting from the left and moving to the right, the first n independent columns {rank % = n). Reorder these columns by taking first Z?i, Ab\, A^bi, etc., until all columns involving bi have been taken; then take Z?2, Ab2, etc., and lastly, take bm, Abm, etc., to obtain % = [bi,Abi,...,Af''-^bi,...,bm,-.^,Af^^-^bml
(4.54)
an n X ^ matrix. The integer ixi denotes the number of columns involving bi in the set of the first n linearly independent columns found in % when moving from left to right. DEFINITION 4.1. The m integers /Xj, / = 1,..., m, are the controllability indices of the system, and ^x = max fxt is called the controllability index of the system. Note that
CHAPTERS:
Controllability, Observability, and Special Forms
284 Linear Systems
E M/
and
(4.55)
mil > n.
To illustrate the significance of ^i, note that an alternative but equivalent definition for IX is that /x is the minimum integer k such that
rank[B,AB,...,A''''B]
(4.56)
= n.
Taking this view, one keeps adding blocks B, AB, A^B, etc., until n independent columns appear (from left to right) for the first time. It is then not difficult to see that /x = max fjLi. Alternatively, jm can be defined as the least integer such that rank [B, AB,...,
A^'^B]
= rank [B, AB,...,
A^B].
(4.57)
Notice that in (4.54) all columns of B are always present since rank B = m. This implies also that /uLi > 1 for all /. Notice further that if A^bi is present, then A^~^bi must also be present in (4.54). For if it were not, i.e., if it were dependent on the previous columns in % and had been eliminated, then A^bt would also have been dependent (write A^'^bi as a linear combination of previous columns in ^ and premultiply by A to show this). Column A^^'bi is of course dependent on the previous ones. This relation can be expressed explicitly as m mmiixuixj)
A^'bt =X
X
y-i
i-\
^iJkA'-'bj + X
k=i
^ij^^'bj
(4.58)
7=1
Note that the first sum in (4.58) indicates the dependence of A^'bi on the linearly independent columns in B, AB,.. .,A^''^B, while the second sum shows the dependence on the independent columns in Af^'B to the left of A^'bi (aijk and pij are appropriate reals). Now define k
a^ = XfJiu
k = l,...,m,
(4.59)
/= i
i.e., cTi == fjLi, cr2 = ^ll + fJi2y..., o"^ = /xi + • • • + ^tm — ^' Also, consider %~^ and let q^, where q^ G R^, k = 1 , . . . , m, denote its akth row, i.e.,
x,qY:-- :X,
^ - 1 = [X,
X, qlV.
(4.60)
Next, define qiA
(4.61)
p^ qmA [qmAf^'i - i
285
It can now be shown that PAP ^ = Ac and PB = B^ with
CHAPTER 3 :
Ac = [Aijl
i,j = 1 , . . . , m,
Controllability, Observability, and Special Forms
0 AH
=
IfJLi-l
Rf^iXf^i^ i = j^
0 X
X
X
0 Aij
R^^^'^J, i # j ,
0 X
and
Be =
X
Bi B2
0
0 J^^.,xm^ (4.62)
Bi = 0
0
1 X
X
where the 1 in the last row of Bi occurs at the /th column location, / = 1 , . . . , m, and X denotes nonfixed entries. Note that Q = CP~^ does not have any particular structure. The expression (4.62) is a very useful form (in control problems) and shall be referred to as the controller form of the system. The derivation of this result is discussed below. First, examples are given to illustrate the procedure involved, and then some alternative expressions and properties are also discussed. EXAMPLE 4.13. Given are A G R'''^'' and B E R""^"^ with (A, B) controllable and with rank B = m. Let n = 4 and m = 2. Then there must be two controllability indices ii\ and 1X2 such that n = 4 = Sf= i i^/ = Mi + M2- Under these conditions, there are three possibilities: (i) fii
Ar =
= 2, /X2 = 2,
All
An
A21
A22.
0
1
:
0
0
X
X
:
X
X Br =
0
0
:
0
1
X
X
:
X
X
Bi] Bi.
(ii) /xi = 1, /X2 = 3, X
Ac =
X
X
X
" 1
0
1
0
0
0
1
X
X
X
Br =
0 0 0
X '
0 0 1
0 1
X
0 0
0 1
0
286
(iii) ixx = 3, /X.2 = 1,
Linear Systems
Ac
X
1
0
0
1
X
X
"0 0 Br = 1
0" 0 X
0
1
X
It is possible to write Ac, Be in a systematic and perhaps more transparent way. In particular, notice that Ac, Be in (4.62) can be expressed as Ac — Ac + t>cAyi where Ac ^ blockdiag
[An, A22,....
B,
—
Jjctjfyi,
(4.63)
Amm\ with
[o An =
//..-I 0 0 0- •0
R^^^'K i =
Be = block diag
\,..,,m
and Am G R^^^ and Bm G R^^^ are some appropriate matrices with ^ ^ ^ ^ /x/ = n. Note that the matrices Ac, Be are completely determined by the m controllability indices iJLi, i = 1 , . . . , m (they are in fact in the so-called Brunovski canonical form— see the discussion following Lemma 4.8). The matrices Am and Bm consist of the (Tith, (T2th,..., cr^th rows of A^ (entries denoted by X) and the same rows ofBc, respectively. Note that Am and Bm is that part of the controllable system x = Ax + Bu that can be altered by linear state feedback u = Fx + Gv, a fact that will be used extensively in the next chapter.
EXAMPLE 4.14. Let A =
1 0 2
0 "0 r 1 and B = 1 1 . To determine the con-1 .0 0.
troller form (4.62), consider =
[B,AB,A^B] 1 0 2
= [bx,h2, Abu Ab2, A^bx, A^b2\ =
where rank^ = 3 = n, i.e., (A, B) is controllable. Searching from left to right, the first three columns of ^ are selected since they are linearly independent. Then
ro ^ = [bi, Abi, Z72] -
lo and the controllability indices are ^ci jjii + iJi2 = 3 = n, and
1 1"
1 0
1
2 OJ
2 and ^2 = 1- Also, ai = /JLI = 2 and (T2 "-1
1
0
0
1
0
1 2 1 2 1 2
Notice that q\ = [0, 0, \] and q2 = [1, 0, - 1 ] , the second and third rows of ^ \ respec-
ro
1 2 1 ,P~' 2 1 2-J
0
tively. In view of (4.61), P = q\A L qi
1 0 1" = 1 1 0 and Ar = .2 0 0.
0 2
PAP~
*c
B{
= PB =
B2\
—
0 1
0 ' 1
0
1
1 One can also verify (4.63) quite easily. We have
Ar + BrAvy
0
and
0 1
o"
0
1
1
0
"o 1
BcBtr
:
0
:
0
:
0
+
o]
0 1
0
0
1J
-1 0
ol 0
ri 1"
[o 1. 0
1J
It is interesting to note that in this example, the given pair (A, B) could have already been in controller form if B were different, but A the same. For example, consider the following three cases:
1. A =
2. A
B =
0
1
0
0
0
1
0
2
-1
0 3. A = 0 0
1 0
0 1 2 - 1
B =
' 1
X
0 0
0 1
0 1
X
0
1
^x
= \, 1X2 = 2 ,
o' /xi = 2, ^1 = 1,
/^i
Note that case 3 is the single-input case (4.42).
•
Several results involving the controllability indices of (A, B) are now presented. First, it is shown that the controllability indices /x/ of a system, defined in Definition 4 . 1 , do not change under similarity transformations [for if they were changing, then (Ac, Be) might have different controllability indices from the original (A, B)].
287 CHAPTER 3 :
Controllability, Observability, and Special Forms
288 Linear Systems
In particular, let x = Ax -\- Buhe given, where (A, B) is controllable, and consider x = Ax -i- Bu, where A = PAP~^, B = PB, x = Px, and P is nonsingular. LEMMA 4.8. The controllability indices of (A, B) are identical to the controllability indices of A = PAP-\ B = PB. Proof. We have ^ = [B, AB,..., A^-^B] = p-^[B, AB,..., A^'^B] = p - ^ ^ . To determine the controllability indices /JLI of (A, B), the first n linearly independent columns of ^ are taken, working from left to right. It is clear that the corresponding n columns of ^ will also be linearly independent, since they are the n columns of ^, each premultiplied by P. Furthermore, as one checks linear dependence of columns in ^, from left to right, if a column is dependent on the previous ones, then the corresponding column in ^ will also be dependent on the previous ones in ^ (show this). Therefore, the linearly independent columns generated by this left to right search are exactly the same in ^ and ^, and therefore, the controllability indices are identical. • Note that the controllability indices of (A, BG), where G is nonsingular, are equal to the controllability indices /x/ of (A, B) within reordering (see Exercise 3.20). Note also that when linear state feedback u = Fx -\- Gv with |G| T^ 0 is applied to X = Ax -^ Bu (see Exercise 3.21), then the controllability indices of (A + BF, BG) are equal to the controllability indices of (A, B), i.e., the controllability indices are invariant under linear state feedback. These results can be summarized as follows. Given (A, B) controllable, then (P(A + BGF)p-\ PBG) will have the same controllability indices, within reordering, for any P, F, and G (\P\ T^ 0, \G\ T^ 0) of appropriate dimensions. In other words, the controllability indices are invariant under similarity and input transformations P and G, and state feedback F [or similarity transformation P and state feedback (F, G)]. Furthermore, it can be shown that if two pairs (A/, Bt), i = \, 2, have the same indices, then there must exist P, G, and F such that Ai = P(A2+ ^2GF)P"^ and 5i = PB2G. This in fdidhXh^ completeness property, which together with the invariance property implies that the controllability indices {/x/} constitute a set of complete invariants for (A, B) under operations P, G, and F, Using (4.63), it is not difficult to see that given any (A2, B2) with controllability indices {/x/}, there exist_P, G, F such that Ai = P(A2 + B2GF)P~^ and Bi = PB2G are exactly equal to Ac = Ai, Be = B\. The pair {Ac, Be) is unique, and since any controllable pair (A, B) with controllability indices {^t/} can be reduced to (Ac, Be) by these operations, {Ac, Be) is a canonical form of such (A, B) under these operations. The pair {Ac, Be) is called the Brunovsky canonical form. It should also be mentioned here that the {jUi} are the right Kronecker indices of the pencil [si - Ac, Be]. (Recall that [si - Ac, Be] = s[I, 0] - [Ac, -Be] = sE - A, a matrix form called a linear matrix pencil.) The derivation of A^, Be in (4.62) is discussed next. The exact structure of Ac and Be depends on the selection of the n linearly independent columns in % and on the choice of the equivalence transformation matrix P. The n linearly independent columns in ^ were selected by a search from left to right; if a column was found dependent on previous ones, then it was dropped. The dependence relation is given in (4.58). This relation, together with the expression for P given in (4.61), are central in the derivation of A^, Be. In view of ^ - ^ ^ = /, qt^ - qi[bi,Ab,,...,Af'^-'bu-.-.A^^~'bi,...,bm.....A^--'b^] - [ 0 , . . . , 0, 1, 0 , . . . , 0], i = l,...,m,
^^^^^
289
where the 1 occurs at the cr^th column. These relations can be written as
CHAPTER 3 :
qtA^'^bi
= 0
^ = 1 , . . . , /x/ - 1,
and
qiA^^-'bi
-
I, (4.65)
where / = 1 , . . . , m, and j = 1 , . . . , m. In view of (4.64), (4.65), and the dependence relations (4.58), it can be shown that P^ is nonsingular, which in view of the fact that 1^1 7^ 0 implies that \P\ T^ 0 as well, i.e., P qualifies as a similarity transformation. We note that the proof of this result is rather involved (see Section 3.7, Notes, for appropriate references). In view of the relationship PA = AcP, it is now not difficult to see that n - m rows of Ac will contain fixed zeros and ones, as in (4.62). The (m) (Tith, c r 2 t h , . . . , cr^th rows of Ac that are denoted by Am in (4.63) are given by qiA^' ^rn
P-\
—
(4.66)
qmAf^^ where Ac = Ac + Be Am. Similarly, in view of the equality Be = PB and (4.63), it is not difficult to see that n - m rows of Be must be zero. The (m) remaining crith, 0-2th,..., amth rows of Be that are denoted by Bm in (4.63) are given by qiA^'-^ B,
Bm —
(4.67)
qmA^^ where Be = BcB The matrix Bm in (4.67) is, in fact, an upper triangular matrix with ones on the diagonal. This result follows from the special form of P^ that was used above to show that P is nonsingular. At this point it may perhaps be beneficial to examine a special case that is not only interesting but will also provide some indication of the type of proof required to establish these results. In particular, it will be shown that when jui < ^l2 — * *' — fjLm, the upper triangular matrix Bm in (4.67) is in fact diagonal. LEMMA 4.9. If/Xi < ^l2
fjLm, then Bm in (4.67) is diagonal with ones on the
diagonal, i.e., Bm = ImProof, The matrix Bm in (4.67) is upper triangular. When, in addition, /xi < )Li2 ^ ••• < fjLm. then qiAf^^-^B - ^iA^i-i[Z?i, Z?2, • •-, W = [1, 0 , . . . , 0], in viewof (4.65); in particular, qiA^^~^bi = 1 and q\Af^^~^b2 = 0 since q\A^~^b2 = 0, k = 1 , . . . , JLXI, and /xi < ^l2. Similar statements hold for the columns bs, b^, ...ybm- The proof for ^2^^2-15^..., qmA^^'^'^B is completely analogous. • Note that if different relations among the controllability indices /x^ exist, then different entries of Bm (above the diagonal) will be zero.
0 EXAMPLE 4.15. We wish to reduce A = 0
0
1 0" 1 r 1 andB = 0 1 to controller 0 2 -1_ _0 0.
form. Note that A and B are almost the same as in Example 4.14; however, presently, 1 < 2 = /X2, as will be seen. We have % = [B, AB, A^B] = [bu bi, Abu Ml
Controllability, Observability, and Special Forms
290 Linear Systems
1 1 0 0 1 0 0 0 0
Abo
1 0 2
. Searching from left to right, the first three linearly in-
1
1 1" 1 0 , from which we 0 0 2.
dependent columns are bi, b2, Ab2 and ^ = [bi, b2, Ab2] = 0
[1 -1 -^1 pute^^i conclude that jLii = 1, fjL2 = 2,ai = 1, and 0-2 = 3. We compute
0 .0
1 0
0 |.
Note that qi = [1, - 1 , - ^] and ^2 = [0, 0, ^], the first and third rows of ^ ~ \ respectively. Then qi ' p =
qi
-
qiA.
"1 - 1 0 0 .0 1
"1 2 r 0 1 1 .0 2 0.
and
•1
0
Ac = PAP-^ 0 A21
c
= PB =
Bi]
52J
=
^22
" 1
0"
0 0
0 1
2
qi B and equals I2, as expected in view of the above qiA lemma (JJLI < 1^2)- In general, Bm is an upper triangular matrix. For the present example, we ask the reader to also verify relations (4.65), determine Am by (4.66), and compare these with the above results. • Presently, (4.67) yields 5^
In the case of single-input systems, m = 1, ^tl = n, and P in (4.61) is exactly the transformation matrix given in (4.44) (verify this). In the case m = 1, if the order of rows in P is reversed and P i in (4.48) is used instead, then an alternative controller form is obtained, shown in (4.47). Similar results apply when m> 1: if the order of the rows of P in (4.61) is reversed within the fit X n blocks, then
qi (4.68)
Pi
in which case Ad
= P\AP^ ^,Bc^ = P\, and B are given by Aci = [Aijl
ij
- 1 , . . . , m,
X
X
291
X
0 Aii
CHAPTERS:
/?M,XM,^ j = j^
=
Controllability, Observability, and Special Forms
V,-i
Aij
X
X
0
0
R^^iX^^j^ i ^ j^
=
0 "fii and
Bi =
Bel =
0 0
••• •••
0
0 0
1 0
X 0
0
0
0
••• •••
X 0
J^t^i
(4.69)
0
where the 1 on the first row of Hi occurs at the iih (i = 1 , . . . , m) location and X denotes nonfixed entries. These formulas are similar to (4.62) and can be derived in a completely analogous fashion. As in the case m = I [see (4.49) to (4.52)], the columns of P~^ can also be selected to be n linearly independent columns of ^ , in which case A = PAP~^ and B = PB will be in alternative controller forms. There are many ways of selecting these n linearly independent columns of the n X nm matrix ^ , as was discussed before. For example, one may search ^ from left to right as above, or one may check bi, Abi,.. .,b2, Ab2,... for linear independence. The controller forms resulting in this way are generahzations of (4.50) and (4.52). These forms are not as useful to us and are omitted here. Structure theorem—controllable version. The transfer function matrix H{s) of the system x = Ax + Bu, y = Cx + Duis given by H(s) = C(sl - Ay^B + D. If (A, B) is in controller form (4.62), then H{s) can alternatively be characterized by the Structure Theorem stated in Theorem 4.10. This result is very useful in the realization of systems, addressed in Chapter 5 and in the study of state feedback in the next chapter. Let A = Ac = Ac + Be Am and B = Be = BcBm, as in (4.63), with | 5 ^ | T^ 0 and let C = Q and D = Dc. Define vf^l e^2
S(s) = block diag
AW^
1 s
/ — 1,..., m
(4.70) Note that S{s) is an n X m polynomial matrix (n = Xf^i i^O. i-^., a matrix with polynomials as entries. Now define the m X m polynomial matrix D(s) and the /? X m polynomial matrix N(s) by D(s) ^ B-'[A(s)
- AmS(s)l
N(s) ^ CcS(s) + DeD(s).
The following is the controllable version of the Structure Theorem.
(4.71)
292
THEOREM 4.10. H(s) = N(s)D~\s),
Linear Systems
Proof, First, note that
where N(s) and D(s) are defined in (4.71).
(si - Ac)S(s) = BcD{s).
(4.72)
To see this, we write BcD{s) = BcBmB;;^^[A(s) - AmS(s)] = BcA(s) - BcAmS(s) and (si - Ac)S(s) = sS(s) - (Ac + BcAm)S(s) = (si - Ac)S(s) - BcAmS(s) = BcA(s) BcAmS(sXwhichpro\ts (4.72). NowH(s) = Cc(sl- AcT^Bc + Dc = CcS(s)D-\s) + Dc = [CcS(s) + DcD(s)]D-\s) = N(s)D-\s). • 0 ro 01 0 and Be = 1 1 , as in Example 4.14. 0. [o 1. 0" r^^ 0 -1 1 11 = r^ 0. y Bm — \ 0 i j .ThenA(5) 0 s 0 '
0 1 EXAMPLE 4.16. Let Ar = 2 - 1 1 0 Here /JLI = 2, iui2 = I and A^ =
2 1
ri 01 S(s) = \s
0 , and
LO i j D(s) = B-'[A(s)-A,nS(s)]
"1.0 Now
IJL
1 .0
=
IJ[LO
's^ + s - I -1
-1
-s + 2 1
S^
0 0
—s s
Cc = [0,1,1], and Dc = [0, 0], N(s) = CcS(s) + DcD(s) = [s, 1],
and
H(s) = [s,l]
5^+5-1 -1
--s
-1
= [s, 1] s [I
S
1 s s^2 + s - l\ s(s^ + 5 - 2 )
= - j ^ - [ 5 ^ + 1,2^^ + 5 - 1 ] s(s^ + 5 - 2 ) = Cc(sl - AcY^Bc + Dc. "0 EXAMPLE 4.17. Let Ac = 0 2
1 0 1
0 1 -2
Be =
Cc =
[0,1,0],
and Dc = 0 (see Example 4.11). In the present case we have A^ = [2,1, - 2 ] , B^ = 1, A(5) = s\S(s) = [l,5,52]^,and D(s) = 1 . [5^ - [2,1, -2][1, 5, s^f]
= 5^ + 2 5 ^ - 5 - 2 ,
N(s) = 5.
Then H(s) =
N(s)D-\s)
+ 25^-5-2
= Cc(sl - AcT^Bc + Dc.
2. Observer forms Consider the system x = Ax + Bu, y = Cx -\- Du given in (4.1) and assume that (A, C) is observable, i.e., rankG = n, where C CA CA' n-\
(4.73)
Also, assume that the p X n matrix C has a full row rank p, i.e.,
293 (4.74)
rankC = p ^ n.
Presently, it is of interest to determine a transformation matrix P so that the equivalent system representation {A^, Bo, Co, Do} with Ao
PAP~\
Bo = PB,
Co = CP~\
Do = D
(4.75)
will have {Ao, Co) in an observer form (defined below). As will become clear in the following, these forms are dual to the controller forms previously discussed and can be derived by taking advantage of this fact. In particular, let A ^ A^ ,B = C^ [(A, B) is controllable] and determine a nonsingular transformation P so that Ac = PAP~^, Be = PB are in controller form given in (4.62). Then Ao = A^ and Co = B^ is in observer form. In fact the equivalent representation in this case is given by (4.75), where P = (P^)~^ (show this). It will be demonstrated in the following how to obtain observer forms directly, in a way that parallels the approach described for controller forms. This is done for the sake of completeness and to define the observability indices. Our presentation will be rather brief. The approach of using duality just given can be used in each case to verify the results. We first note that if rank C = r < /?, an approach analogous to the case when rank B < m can be followed, as above. The fact that the rows of C are not linearly independent means that the same information can be extracted from only r outputs, and therefore, the choice for the outputs should perhaps be reconsidered. Now if (A, C) is unobservable, one may use two steps to first isolate the observable part and then reduce it to the observer form, in an analogous way to the uncontrollable case previously given. Single-output case (p = I). Let p - i = (2 A
[q^Aq,...,A''-^q],
(4.76)
where q is the ^th column in 0~^ Then 0 0
-ao -ai
An =
Co = [0,...,0,1L 1
(4.77)
-OCn-\
where the at denote the coefficients of the characteristic polynomial a(s) = det(sI-A) = ^^ + a:„-i^'^"i + - - - + a i ^ + ao.HereA^ = PAP'^ = Q~^AQXo = CP' ^ = CQ, and the desired result can be established by using a proof that is completely analogous to the proof in determining the (dual) controller form presented earlier in this section. Note that Bo = PB does not have any particular structure. The representation {Ao, Bo, Co, Do} will be referred to as the observer form of the system. Reversing the order of columns in P~ ^ given in (4.76) or selecting P to be exactly 0, or to be equal to the matrix obtained after the order of the columns in 0 has been reversed, leads to alternative observer forms in a manner analogous to the controller form case. We leave it to the reader to investigate these possibilities further.
CHAPTERS:
Controllability, Observability, and Special Forms
294 Linear Systems
EXAMPLE 4.18. Let A
-1 0 0
0 0 1 0 and C = [1, -1,1]. To derive the ob0 - 2
server form (4.77), we could use duality, by defining A = A^, B = C^, and deriving the controller form of A, B, i.e., by following the procedure outlined above. This is left to the reader as an exercise. We note that the A, B are exactly the matrices given in Examples 4.11 and 4.12. In the following the observer form is derived directly. In particular, we have
c CA CA\
" 1 - 1 = -1 -1 . 1 - 1
1" -2 , 4.
1
- 21
1 2
1 3
1 2
1 6
.-1
0
1 3
0-1 =
.76), and in view of (4.76), r
1
Q = p - i = [q^Aq^A^q] = L
Note that q = Ao = Q-'AQ
I I 6' 3
2
1 2
1 1 2
1 6
1 6
1 6
1 3
2
4
3
3 J
, the last column of 0"^. Then 0 0 1
=
2 1 -2
and
Co = CQ = [0, 0,1],
where l^-/ - A| = s^ + 2s - s - 2 = s^ -{- a2S^ + ais + ao. Hence, ii 2
i
GA. =
6
= AQ.
Multi-output case (p > 1). Consider ci
C CA
ciA CpA
(4.78)
CA « - i ciA « - i LCpA'^-
where c\,.. .,Cp denote thep rows of C, and select the first n linearly independent rows in 0, moving from the top to bottom (rank € = n). Next, reorder the selected rows by first taking all rows involving ci, then C2, etc., to obtain
295 c\A ciA
CHAPTERS:
v^-l
(4.79)
Sin n X n matrix. The integer Vi denotes the number of rows involving c/ in the set of the first n Unearly independent rows found in 0 when moving from top to bottom. DEFINITION 4.2. The/? integers vt, i = 1,..., p, are the observability indices of the system, and v = max Vi is called the observability index of the system. Note that Y^Vi = n
and
(4.80)
pv ^ n.
When rankC = p, then ^/ > 1. Now define k
d-k =^Pi
k = I,
(4.81)
A
/-I
i.e., (Ti = 1^1, (T2 = v\ -}- V2y' "y^p = v\ + '" -\- Vp = n. Consider 0 ^ and let qk E R^, k = 1 , . . . , /?, represent its or^th column, i.e., g - i = [X--- X^il X ••• Xg2|**-| X ••• X qpl
(4.82)
Define P-' ThenA^ = PAP'^
= Q= [qi,...,A-^-'q,,...,qp,.,,,A^p-'qpl = Q-^AQmdCo
= CP~^ = CQ are given by
Ao = [Aijl i,j=\,..., 0
•••
0
(4.83)
p,
X
/^^^•^^/ = 7,
An =
ro ••
Co - [Cl, C2,..., Cp], Ci —
X
0
X
Aij =
/^^'•><^^/#y,
X
and
0
0
0
0 0 0
0 1 X
0
X
J^p-XVi
(4.84)
where the 1 on the last column of C/ occurs at the /th-row location (/ = \,..., p) and X denotes nonfixed entries. Note that the matrix BQ = PB = Q~^B does not
Controllability, Observability, and Special Forms
296 Linear Systems
have any particular structure. Equation (4.84) is a very useful form (in the observer problem) and shall be referred to as the observer form of the system. Analogous to (4.63), we express AQ and Co as /\.o
^o
' ^p^O)
^o
(4.85)
^p^oy
0
•• RVi^^Vi^ C^
where Ao = blockdiag [A\, A2,..., Ap\ with A/ =
=
0 block diag {[0,...,0,lY e 7?^s/ = 1 , . . . , ;?), and Ap G T?"^^, and Cp G RP'^P are appropriate matrices (Xf= 1 ^/ = ^). Note that AQ, CO are completely determined by the/? observability indices ^'/, / = I,..., p, and A^ and Cp contain this information in the o-ith,..., or^th columns of Ao and in the same columns of Co, respectively. Note also that Ap is that part of the observable system i = Ax + Bu, y = Cx^- Du that can be altered by the gain of a state observer. This is discussed further in the next chapter. Results that are in the spirit of Lemmas 4.8 and 4.9 also exist and can be established directly in a completely analogous way or by duality. The same is true for proving that Q in (4.83) is nonsingular and that Ao and Co in (4.84) have their particular structure. Furthermore, by reversing the order of the columns of P~^ in (4.83) or by selecting the columns of P directly from 0, one can derive alternative observer forms. These are dual to the controller forms discussed before. ro EXAMPLE 4.19. Given A = 1
0
01
0
2 andC =
"0 1
1 0" we wish to reduce 1 0_
LO 1 -ij these to observer form. This can be accomplished using duality, i.e., by first reducing A = A^, 5 = C^ to controller form. Note that A, B are the matrices used in Example 4.14, and therefore, the desired answer is easily obtained. Presently, we shall follow the direct algorithm described above. We have 0 1 1 1 0 0
C CA CA^
1 1 0 0 2 2
0 0 2 2 -2 -2
Searching from top to bottom, the first three linearly independent rows are ci, C2, c\A, and "0 1 .1
' ci '
c\A
=
. C2 .
1 0" 0 2 1 0.
Note that the observability indices are z^i = 2, z^2 = lando-i == 2,0-2 = 3. We compute
Then, Q = [qi, Aqi, ^2]
0 =
r
"X X
0 0
1 2J
L^
2
T andg-i = 0 .1
1" 0 2^
1 2" 1 0 . Therefore, 0 0.
0 1
Ao = Q-'AQ = All Co = CQ = [Ci
All,
: Ci\ = 0 0
297
2 -1
CHAPTERS:
Controllability, Observability, and Special Forms
0
1 : 0 1 : 1
We can also verify (4.85), namely,
Ao
0
2
1
-1
0
0
— Ao + ApC^o
0
0
0
1
0
0
0
0
0
" 2 11 ro 1 0' + -1 0 [o 0 1. 0 Oj
and Co =
0
1 :
0
1 :
^ p^o
1 0 0 1 0 1 1 0 0 1
Structure Theorem—observable version. The transfer function matrix H{s) of
systemi = Ax + Bu,y = Cx + Duisgiytnhy H{s) =
C(sI-A)~^B-^D,lf(A,C)
is in the observer form, given in (4.84), then H(s) can alternatively be characterized by the Structure Theorem stated in Theorem 4.11. This result will be very useful in the realization of systems, addressed in Chapter 5 and also in the study of observers in the next chapter. Let A = Ao = Ao + ApCo and C = Co = CpCo as in (4.85) v^ith \Cp\ # 0, let B = Bo and D = Do, and define A(^) = diag [^^S s''^ . . . , s^'p],
S(s) = block diag ([1, ^ , . . . , s""^'^], / = 1,..., p). (4.86)
Note that S{s) is a p X n polynomial matrix, where n = ^f^^vi. Now define the pX p polynomial matrix D(s) and the p X m polynomial matrix N(s) as D(s) ^ [Ms) - S(s)Ap]Cp
N(s) = S(s)Bo + D(s)Do.
(4.87)
The following result is the observable version of the Structure Theorem. It is the dual of Theorem 4.10 and can therefore be proved using duality arguments. The proof given is direct. THEOREM4.il. H{s) = D~H^)^('y), where 7V(5), 6(5) are defined in (4.87). Proof, First we note that b{s)Co = S(s){sl - Aol
(4.88)
To see this, write D(5)Co = [A(s)-S(s)Ap]C-^CpCo = A(5)C^-5^(>y)A^Co, and also, S(s)(sl - Ao) = S(s)s - S(s)(Ao + ApCo) = S(s)(sl - Ao) - S(s)ApCo = A(^)C. S(s)ApCo, which proves (4.88). We now obtain H{s) = Co{sI - Ao)~'^Bo + Do = D-\s)S(s)Bo + Do = D~\s)[S(s)Bo + D(s)Do] = D-\s)N(s), •
298 Linear Systems
EXAMPLE 4.20. Consider A^ =
4.19. Here z^i = 2, 1^2 = I, Ms) =
[A(s) -s + 2 0
S(s)Ap]C-' i 0
1 0 -1 1
0
s
0 1 0
2 -1 0 s^ 0
1 0
and Co =
0 0
1 1
0 1 , and S{s) = s 0
s 0
0 . Then D{s) = 1
5 0 0 1
s"- + s-- 2 - 1 0
2 -1 0
1 0 0.
1 0 -1 1
1 1
0 of Example 1
01 - 1 1 's'^ + s - \
's^ 0
01 s -1
Now if Bo = [0, 1, 1]^, Do = 0, and A^(^) = S(s)Bo + D(s)Do = [s, if, then H(s) = D-\s)N(s) = {l/[s(s^ + s- 2)]}[s^ + h2s^ + s- If = Co(sI - AoT^Bo + Do. •
3.5 POLES AND ZEROS In this section the poles and zeros of a time-invariant system are defined and discussed. The primary reason for considering poles and zeros of a system at this time is that poles and zeros are related to the (controllable and observable, resp., uncontrollable and unobservable) eigenvalues of A. These relationships shed light on the eigenvalue cancellation mechanisms encountered when input-output relations, such as transfer functions, are formed. These relationships also provide greater insight into the realization theory addressed later in this book. Furthermore, the relationships between uncontrollable or unobservable eigenvalues, decoupling zeros, and cancellations in the transfer function matrix to be discussed below, provide also insight into how state feedback can alter system behavior. State feedback will be studied later in Chapter 4. In the following development, the fmio^ poles of a transfer function matrix H(s) [or H(z)] are defined first (for the definition of poles at infinity, refer to Exercise 3.36). It should be noted here that the eigenvalues of A are sometimes called poles of the system {A, B, C, D}. To avoid confusion, we shall use the complete itrxn. poles of H(s), when necessary. The zeros of a system are defined using internal descriptions (state-space representations). Smith and Smith-McMillan forms To define the poles of H(s), we shall first introduce the Smith form of a polynomial matrix P(s) and the Smith-McMillan form of a rational matrix H(s). The Smith form Sp(s) of a p X m polynomial matrix P(s) (in which the entries are polynomials in s) is defined as (see also Subsection 7.2C of Chapter 7) Sp(s) =
'A(s) 0
(5.1)
with A(^) = diag [ei(s),..., er(s)], where r = rank P(s). The unique monic polynomials €i(s) (polynomials with leading coefficient equal to one) are the invariant factors of P(s). It can be shown that ei(s) divides e/+i(5'), / = 1 , . . . , r - 1. Note that
€i(s) can be determined by
299
ei(s) =
Di(s)
CHAPTERS:
i = 1, . . . , r ,
where Di(s) is the monic greatest common divisor of all the nonzero /th-order minors of P(s) with Do(s) = 1. The Di(s) are the determinantal divisors of ^(5"). A matrix P{s) can be reduced to Smith form by elementary row and column operations (see Subsections 7.2B and C in Chapter 7). This ensures that the properties of interest are preserved when P{s) is reduced to its (unique) Smith form. In particular, we are interested here in the invariant factors ei{s) of P{s) that can be determined directly from the determinantal divisors Di{s) without having to reduce P{s) to its Smith form via elementary operations. Consider now a /? X m rational matrix H{s). Let d{s) be the monic least common denominator of all nonzero entries and write 1
H{s)
Ks)
N{s),
(5.2)
where N{s) is a polynomial matrix. Let SN{S) = diag [ni(s\ . . . , nr(s), Op-r,m-r] be the Smith form of N(s), where r = rank N(s) = rank H(s). Divide each nds) of SN(S) by d(s), cancelling all common factors to obtain the Smith-McMillan form of H(s), SMH(S)
=
'A(s) 0
0 0
(5.3)
with A(s) = diag [e\(s)/ilji(s),.. .,er(s)/il/r(s)], where r = rank H(s). Note that ei(s) divides ei+i(s), i = \,2,.. .,r ~ I, and if/i+iis) divides iptis), i = \,2,..., r- 1. Poles Given a. pX m rational matrix H(s), its characteristic polynomial or pole polynomial, PH(S), is defined as PH(S) =
il/i(sy"il/r(s\
(5.4)
where the i///, / = 1 , . . . , r, are the denominators of the Smith-McMillan form of H(s). It can be shown that PH(S) is the monic least common denominator of all nonzero minors of H(s). DEFINITION 5.1. The poles ofH(s) are the roots of the pole polynomial PH(S).
•
Note that the monic least common denominator of all nonzero first-order minors (entries) of H(s) is called the minimal polynomial of H(s) and is denoted by mnis). The mH(s) divides PH(S) and when the roots of PH(S) [poles of H(s)] are distinct, J^H(S) = PH(S), since the additional roots in PH(S) are repeated roots of mnis). It is important to note that when the minors of H(s) [of order 1,2, . . . , min (p, m)] are formed by taking the determinants of all square submatrices of dimension 1 x 1 , 2 x 2 , etc., all cancellations of common factors between numerator and denominator polynomials should be carried out. In the scalar case, p = m = 1, Definition 5.1 reduces to the well-known definition of poles of a transfer function H(s), since in this case there is only one
Controllability, Observability, and Special Forms
300
Linear Systems
minor (of order 1), H(s), and the poles are the roots of the denominator polynomial of H{s). Notice that in the present case it is assumed that all the possible cancellations have taken place in the transfer function of a system. Here PH(S) = mnis), that is, the pole or characteristic polynomial equals the minimal polynomial of H(s). Thus, PH(S) = ^H(S) are equal to the (monic) denominator of 77(5'). [l/[^(^ + 1)] 1/^ 1 1 ^, . ^ , EXAMPLE 5.1. Let H(s) ' ^ ^ ^ / 9 • The nonzero minors of order 0 0 l/s^j 1 are the nonzero entries. The least common denominator is 5'^(^ + 1) = mnis), the minimal polynomial of H(s). The nonzero minors of order 2 are l/[s^(s +1)] and 1/^-^ (taking columns 1 and 3, and 2 and 3, respectively). The least common denominator of all minors (of order 1 and 2) is s^(s + 1) = PH(S), the characteristic polynomial of H(s). The poles are {0, 0, 0, -1}. Note that mnis) is a factor of PH(S) and the additional root at 5 = 0 in PH(S) is a repeated pole. To obtain the Smith-McMillan form of H(s), write 1 s(s + 1) s\s + 1) 1 H(s) N(sl (s + 1) sHs + 1) 0 d(i) where d(s) = s^(s + 1) = fnnis) [see (5.2)]. The Smith form of N(s) is SN(S) =
1 0 0 0 s{s + 1) 0
since Do = I, Di == I, D2 = s{s -\- I) [the determinantal divisors of A^(5)], and ni = DJDQ = l,n2 = D2/D1 = s(s + 1), the invariant factors of N(s). Dividing by d(s), we obtain the Smith-McMillan form of H(s), el SMH(S)
=
h
0
0
51 i//2
1 sHs + 1) 0
0
1
0 0
Note that 1/^2 divides ij/i and ei divides €2. Now the characteristic or pole polynomial of H{s) is PH(S) = il^iil^i = s^(s -\- 1) and the poles are {0, 0, 0,-1}, as expected. • 1 EXAMPLE 5.2. Let H(s) = If a 7^ 1, then the second-order minor is s+2 1^(^)1 = (1 - a)/(s + 2)^. The least common denominator of this nonzero second-order minor \H(s)\ and of all the entries of H(s) (the first-order minors) is (s + 2)^ = PH(S), i.e., the poles are at {-2, -2}. Also, mnis) = s + 2. Now if a = 1, then there are only first-order nonzero minors (\H(s)\ = 0). In this case PH(S) = mnis) = S + 2, which is quite different from the case when a 7^ 1. Presently, there is only one pole at - 2 . The reader should verify these results, using the Smith-McMillan form of H(s). m In view of Subsection 3.4.C, it is clear that all the poles of H(s) are roots of l^/ - All |, that is, they are some or all of the controllable and observable eigenvalues of the system. In fact, as will be shown in Chapter 5, the poles of H(s) are exactly the controllable and observable eigenvalues of the system (in An) and no factors of l^-/ - All I in H(s) cancel. In general, for the set of poles of H(s) and the eigenvalues of A, we have {poles of H(s)} C {eigenvalues of A}
(5.5)
with equality holding when all the eigenvalues of A are controllable and observable eigenvalues of the system. Similar results hold for discrete-time systems and H(z).
EXAMPLE 5.3. Consider A
ro - 1 1 -2
[o
1 0" 11 1 ,B = 1 1 and C = [0, 1, 0] (refer
1 -ij
.1
2.
to Example 4.6 in Section 3.5). Then the transfer function H(s) = [l/s, l/s]. H(s) has only one pole, ^i = 0(PH(S) = s), and Ai = 0, is the only controllable and observable eigenvalue. The other two eigenvalues of A, A2 = - 1 , A3 = - 2 , that are not both controllable and observable, do not appear as poles of H(s). • EXAMPLE 5.4. Recall the circuit in Example 4.9 in Section 3.4. If R1R2C ¥^ L, then {poles of//(5')} = {eigenvalues of A at Ai = -\I{R\C) and A2 = -R^II^. In this case, both eigenvalues are controllable and observable. Now if R1R2C = L with Ri 7^ R2, then H(s) has only one pole, ^i = -R2IL, since in this case only one eigenvalue Ai = -R2/L is controllable and observable. The other eigenvalue A2 at the same location -R2IL is uncontrollable and unobservable. Now if R\R2C = L with Ri = R2 = R, then one of the eigenvalues becomes uncontrollable and the other (also at -RIL) becomes unobservable. In this case H{s) has nofinitepoles {H{s) = \IR). • Zeros In a scalar transfer function H{s) the roots of the denominator polynomial are the poles, and the roots of its numerator polynomial are the zeros of H{s). As was discussed, iht poles ofH{s) are some or all of the eigenvalues of A (the eigenvalues of A are sometimes also cdlXtd poles of the system {A, B, C, D}). In particular, it v^as shown in Subsection 3.4.C that the uncontrollable and/or unobservable eigenvalues of A can never be poles of H(s). In Chapter 5 it is shown that only those eigenvalues of A that are both controllable and observable appear as poles of the transfer function H(s). Along similar lines, the zeros ofH(s) (to be defined later) are some or all of the characteristic values of another matrix, the system matrix P(s). These characteristic values are called the zeros of the system {A, B, C, D]. The zeros of a system for both the continuous- and discrete-time case are defined and discussed next. We consider now only finite zeros. For the case of zeros at infinity, refer to the exercises. Let the system matrix (also called Rosenbrock's system matrix) of {A, B, C, D] be P{s)^
si - A -C
Note that in view of the system equations x P(s)
-x(s) U(s)
B D Ax + Bu, y
(5.6) Cx + Du, we have
0
where x(s) denotes the Laplace transform of x(t). Let r = rank P(s) [note that n ^ r ^ min (p-\-n,m-\- n)] and consider all those rth-order nonzero minors of P(s) that are formed by taking the first n rows and n columns of P(s), i.e., all rows and columns of 5'/ - A, and then adding appropriate r - n rows (of [-C, D]) and columns (of [B^, D^Y). The zero polynomial of the system {A, B, C, Z)}, zp{s), is defined as the monic greatest common divisor of all these minors. DEFINITION5.2. The zeros of the system {A, B, C, D} or the system zeros are the roots of the zero polynomial of the system, zp(s). •
301 CHAPTER 3 :
Controllability, Observability, and Special Forms
302 Linear Systems
In addition, we define tlie invariant zeros of the system as tlie roots of tlie invariant polynomials of P{s). In particular, consider the (p + n) x (m + n) system matrix P{s) and let Sp{s)
'A{s) 0
0" 0
A(^) = diag[ei (^),..., e^(^)],
(5.7)
be its Smith form. The invariant zero polynomial of the system {A, 5, C, D} is defined as Z^p{s) = ei{s)e2{s)---er{s) (5.8) and its roots are the invariant zeros of the system. It can be shown that the monic greatest common divisor of all the highest order nonzero minors of P(^) equals Zp{s). In general, {zeros of the system} D {invariant zeros of the system}. When p = m with det P{s) ^ 0, then the zeros of the system coincide with the invariant zeros. Now consider the nx {m-\-n) matrix [si — A,B] and determine its n invariant factors e/(^) and its Smith form. The product of its invariant factors is a polynomial, the roots of which are the input-decoupling zeros of the system {A,5,C,D}. Note that this polynomial equals the monic greatest common divisor of all the highest order nonzero minors (of order n) of [si — A,B]. Similarly, consider the {p-\-n) xn \sl — A\ matrix ^ and its invariant polynomials, the roots of which define the outputdecoupling zeros of the system {A,5,C,D}. Using the above definitions it is not difficult to show that the input-decoupling zeros of the system are eigenvalues of A and also zeros of the system {A,5,C,D} (show this). In addition note that if A/ is such an input-decoupling zero, then rank [A// —A,5] < n, and therefore, there exists a 1 x n vector v/ ^ 0 such that v/[A// —A,5] = 0. This, however, implies that A/ is an uncontrollable eigenvalue of A (and Vi is the corresponding left eigenvector), in view of Subsection 3.4B. Conversely, it can be shown that an uncontrollable eigenvalue is an input-decoupling zero. Therefore, the input-decoupling zeros of the system {A,5,C,D} are the uncontrollable eigenvalues of A. Similarly, it can be shown that the output-decoupling zeros of the system {A,5,C,D} are the unobservable eigenvalues of A. They are also zeros of the system, as can easily be seen from the definitions. There are eigenvalues of A that are both uncontrollable and unobservable. These can be determined using the left and right corresponding eigenvector test or by the Canonical Structure Theorem (Kalman Decomposition Theorem) (see Subsections 3.4A and B). These uncontrollable and unobservable eigenvalues of A are zeros of the system that are both input- and output-decoupling zeros and are called input-output decoupling zeros. These input-output decoupling zeros can also be defined directly from P{s) given in (5.6); however, care should be taken in the case of repeated zeros. If the zeros of a system are determined and the zeros that are input- and/or output-decoupling zeros are removed, then the zeros that remain are the zeros of H{s) and can be found directly from the transfer function II{s). In particular, if the Smith-McMillan form of H{s) is given by (5.3), then ZH{S) =ei{s)e2{s)"'er{s)
(5.9)
is the zero polynomial of H(s) and its roots are the zeros of H(s). These are also called the transmission zeros of the system. DEFINITION 5.3. The zeros of H(s) or the transmission zeros of the system are the roots of the zero polynomial of H(s), ZH(S)' • The relationship between the zeros of the system and the zeros of H(s) can easily be determined using the identity P(s) =
si
-A -C
si
B D
-A -C
(si - A)- 'B H(s)
for the case when P(s) is square and nonsingular. Note that in the present case |P(^)| = 1^/ - A||//(5')|. It is also possible to obtain this result using the special structure of the matrices in the special form of Subsection 3.4A. In this case, the invariant zeros of the system [the roots of IPC^")]], which are equal here to the zeros of the system, are the zeros of H(s) [the roots of |//(^)|] and those eigenvalues of A that are not both controllable and observable [the ones that do not cancel in \sl - A\\H(s)\], Note that the zero polynomial of H(s), ZH(S), equals the monic greatest common divisor of the numerators of all the highest order nonzero minors in H(s) after all their denominators have been set equal to PH(S), the characteristic polynomial of H{s). In the scalar case (p = m = I) our definition of the zeros of H{s) reduces to the well-known definition of zeros, namely, the roots of the numerator polynomial of H(s). EXAMPLE 5.5. Consider H(s) of Example 5.1. From the Smith-McMillan form of H(s), we obtain the zero polynomial ZH(S) = 1, and H(s) has no (finite) zeros. Alternatively, the highest order nonzero minors are l/[s^(s + 1)] and l/s^ = (s + l)/[s^(s +1)] and the greatest common divisor of the numerators is ZH(S) = 1. r s 0 s+ 1 EXAMPLE 5.6. We wish to determine the zeros of H(s) = s+ 1 1 s+ 1 s+ 1 1 The first-order minors are the entries of H(s), namely, and s + I' s + I' s^ s+ 1 1 Then PH(S) = s^(s + 1), there is only one second-order minor s+I s^ s' the least common denominator, is the characteristic polynomial. Next, write the highest s(s + 1) _ s(s + 1) and note that s(s + 1) is the (second-) order minor as s sKs + 1) " PH(S) zero polynomial of H(s), ZH(S), and the zeros of H(s) are {0, -1}. It is worth noting that the poles and zeros of H(s) are at the same locations. This may happen only when H(s) is a matrix. If the Smith-McMillan form of H(s) is to be used, write H(s) 's'
0
U'
(s + If
sHs + 1) 0 + 1)2 since
—--N(s). The Smith form of N(s) is now S\S d(s) Do = l,Di -- l,D2 = ^-^(^+1)2 with invariant factors of A/^(5') given by ^1 = DI/DQ = 1 and ^2 = D2/D\ s^(s+lf. Therefore, the Smith-McMillan form (5.3) of H{s) is SMH(S)
=
1 s\s + 1) 0
r^l
0
01
^l
s(s + 1)
i
J
0
^2 IA2-I
303 CHAPTER 3 :
Controllability, Observability, and Special Forms
304
Linear Systems
The zero polynomial is then ZH(S) = e\e2 = s(s + 1) and the zeros of H(s) are {0, -1}, as expected. Also, the pole polynomial is PH(S) = i//ii/^2 = s^{s + 1) and the poles are {0,0,-1}. •
EXAMPLE 5.7. We wish to determine the zeros of H(s)
s+ 1 1 s+ 1
0
5+ 1
The
1 s and the characteristic polynomial is PH(S) 0
1 1 1 s' s + I' s(s + I) s^(s + 1). Rewriting the highest (second-) order minors as s(s + \)IPH{S), s^lpnis), and SIPH{S), the greatest common divisor of the numerators is s, i.e., the zero polynomial of H{s) is ZH{.S) = s. Thus, there is only one zero of H(s) located at 0. Alternatively, note that the Smith-McMillan form is 1 sHs + 1) SMH(S) 0 second-order minors are
0
Relations between poles, zeros, and eigenvalues of A Consider the system x = Ax -\- Bu, y = Cx -\- Du and its transfer function matrix H(s) = C(sl -A)~^B-\-D. Summarizing the above discussion, the following relations can be shown to be true. 1. We have the set relationship {zeros of the system} = {zeros of H{s)} U {input-decoupling zeros} U {output-decoupling zeros} - {input-output decoupling zeros}. (5.10) Note that the invariant zeros of the system contain all the zeros of H(s) (transmission zeros), but not all the decoupling zeros (see Example 5.8). When P(s) is square and nonsingular, the zeros of the system are exactly the invariant zeros of the system. Also, in the case when {A, B, C, D} is controllable and observable, the zeros of the system, the invariant zeros, and the transmission zeros [zeros of H(s)] all coincide. 2. We have the set relationship {eigenvalues of A (or poles of the system)} = {poles of H(s)} U {uncontrollable eigenvalues of A} U {unobservable eigenvalues of A} - {both uncontrollable and unobservable eigenvalues of A}. (5.11) 3. We have the set relationships
and
{input-decoupling zeros} = {uncontrollable eigenvalues of A}, {output-decoupling zeros} = {unobservable eigenvalue of A}, {input-output decoupling zeros} = {eigenvalues of A that are both uncontrollable and unobservable}. (5.12)
305
4. When the system {A, B, C, D} is controllable and observable, then
CHAPTER 3 :
{zeros of the system} = {zeros of H(s)} and
{eigenvalues of A (or poles of the system)} = {poles of H(s)}.
(5.13)
Note that the eigenvalues of A (the poles of the system) can be defined as the roots of the invariant factors of ^/ - A in P(s) given in (5.6). EXAMPLE 5.8. Consider the system {A, B, C} of Example 5.3. Let s P(s) =
si - A -C
-1
B D
-1
1
0
^+ 2
-1
1
1
-1
s+ 1
1
2
0
0
0
1
0
0
-1
There are two fourth-order minors that include all columns of ^/ - A obtained by taking columns 1, 2, 3, 4 and columns 1, 2, 3, 5 of P{s)\ they are {s + \){s + 2) and {s + \){s + 2) (verify this). The zero polynomial of the system is zp = {s+ l){s + 2) and the zeros of the system are {-1, - 2 } . To determine the input-decoupling zeros, consider all the third-order minors of [si - A, 5 ] . The greatest common divisor is 5* + 2 (verify this), \sl — Al which implies that the input-decoupling zeros are {-2}. Similarly, consider and show that 5 -t- 1 is the greatest common divisor of all the third-order minors and that the output-decoupling zeros are {-1}. The transfer function for this example was found in Example5.3 tobe//(>y) = [I/5, 1/^]. The zero polynomial of//(i-) is z//(^) = land there are no zeros of H{s). Notice that there are no input-output decoupling zeros. It is now clear that relation (5.10) holds. The controllable (resp., uncontrollable) and the observable (resp., unobservable) eigenvalues of A (poles of the system) have been found in Examples 4.4 and 4.5 in Section 3.4. Compare these results to show that (5.12) holds. The poles of H(s) are {0}. Verify that (5.11) holds. One could work with the Smith form of the matrices of interest and the SmithMcMillan form of H(s). In particular, it can be shown (do so) that the Smith form of P(s) is
"1
0
0 0
1 0
"1 0 0 0
0 1 0 0 0
«
'^ 0 0 s+2
0^ ^1 I, of [si - A, B] is '
no
0 , and of [si -- A ] is 0 s+1 0 0 0 0 the Smith-McMillan form of H(s) is
0 0 s+2
0
01 of
si
-A -C
0
1 0 0 ^r^ -1-1){
SMH(S)
0
=
. Also, it can be shown that
-,o
It is straightforward to verify the above results. Note that in the present case the invariant zero polynomial is Zp(s) = s + 2 and there is only one invariant zero at - 2 . • EXAMPLES.9. Consider the circuit of Example 5.4 and of Example 4.9 in this chapter and the system matrix P(s) for the case when R1R2C = L given by
Controllability, Observability, and Special Forms
306 Linear Systems P(s) =
si -A -C
B D
s + R2 L
0
0
s + R2
Ri
^
R2
L
1 '
Ri
(i) First, let Ri ¥^ R2. To determine the zeros of the system, consider \P{s)\ = {\IR\){s + R\IL){s + R2IL), which impHes that the zeros of the system are {—R\IL, -R2IL}. Consider now all second-order (nonzero) minors of [si - A,B], namely, {s + R2/Lf. (l/LXs + R2/L), and -(R2/L)(s + R2/LX from which we see that {~R2/L} is the input-decoupling zero. Similarly, we also see that {-7^2/^} is the output-decoupling zero. Therefore, {-R2/L} is the input-output decoupling zero. Compare this with the results in Example 5.4 to verify (5.13). (ii) When Ri = R2 = R, then \P(s)\ = (l/R)(s + RILf, which impHes that the zeros of the system are at {-R/L, -R/L}. Proceeding as in (i), it can readily be shown that {—R/L} is the input-decoupling zero and {—R/L} is the output-decoupling zero. To determine which are the input-output decoupling zeros, one needs additional information to the zero location. This information can be provided by the left and right eigenvectors of the two zeros at -R/L to determine that there is no input-output decoupling zero in this case (see Example 4.9). In both cases (i) and (ii), H(s) has been derived in Example 4.9 of Section 3.4. Verify relation (5.10). • Zero directions and pole-zero cancellations There are characteristic vectors or zero directions, associated with each invariant and decoupling zero of the system {A, B, C, Z)}, just as there are characteristic vectors or eigenvectors, associated with each eigenvalue of A (pole of the system). Consider the system matrix P(A) given in (5.6) at 5* = A. If A is an invariant zero of the system, then there exist nonzero vectors v^ and v^ associated with A such that P(A)v, = 0,
v,P(A) = 0.
(5.14)
This is so because if A is an invariant zero of the system, then ranii P(X) < rank P(s) < min(^ -\- p,n -\- m), and therefore the columns (resp., the rows) are linearly dependent. Here rank P(s) denotes the number of linearly independent columns or rows over the field of rational functions in s [it is called normal rank of P(s)]; rank P(A) is the rank of P(A) over the field of complex numbers. The vector v^ is called an invariant zero direction or a zero direction corresponding to A. A physical interpretation of this is as follows. Let v^ = -u , that is. let XI - A B (5.15) -C D r..,^t and assume that the system is at rest at ^ = 0. If an input of the form w(0 = ue 0, is applied, and if A is not a pole of the system, then it will produce a state of the form x{t) = xe^^ and an output
yit) - 0,
r > 0.
This is sometimes referred to in the literature as the output-zeroing or blocking property of zeros. This property is a generahzation of the blocking property of zeros originally expressed in terms of the transfer function and its zeros for a SISO system. For decoupling zeros, it is quite easy to see what the corresponding directions will be. In particular, if A is an input-decoupling zero, then there exists a nonzero vector V such that v[A/ - A, B] = 0, which implies that v,PW
= [v,0]
- A B ]
\XI
-c
D\
[0,0].
(5.16)
It is clear that the vector v is the left eigenvector of A corresponding to A, which is also an eigenvalue of A. This, in fact, determines the exact relationship (location and corresponding directions) between input-decoupling zeros and uncontrollable eigenvalues. Similar results can be derived for the output-decoupling zeros. In fact if A is an [A/ — AI v-0, output-decoupling zero, then there exists a nonzero vector v so that -C which implies that ^(A)v, =
XI -A -C
(5.17)
It is clear that the vector v is the right eigenvector of A corresponding to A, which is also an eigenvalue of A. This provides the exact relation between output-decoupling zeros and unobservable eigenvalues. Consider now an input-decoupling zero A and the corresponding direction [v, 0]. As was shown above A, v are an (uncontrollable) eigenvalue and its corresponding left eigenvector, respectively. Let v be the corresponding right eigenvector to A. If \I_ — A] ^ V = 0, then A is also an unobservable eigenvalue and an output-decoupling zero. In particular, A is an input-output decoupling zero and also an eigenvalue that is both uncontrollable and unobservable. The vectors [v, 0] and [v^, 0]^ are the directions associated with such zero. In general the directions v^ and v^ in (5.14) associated with the invariant zeros do not have any particular form [v^ = [v, 0] in (5.16) for the case of an input-decoupling zero]. When the rank of P(s) is full, that is, rank P(s) = n -{- min (p, m),
(5.18)
the direction that corresponds to the invariant zero at A can be taken to be v^ [where P{X)Vz = 0] when min (;?, m) = m, andv^ [where V2P(A) = 0], when min (/?, m) = p. Note that when min (p, m) = p < m, there are nonzero vectors satisfying P{K)Vz = 0 with A not necessarily being an invariant zero of the system. This situation becomes accentuated, i.e., (5.14) is satisfied for values of A that are not necessarily zeros of the system, when rank P(s) < n-\-mm {p, m). This phenomenon is unique to MIMO systems. In the MIMO case it is possible for a /? X m transfer function H{s) to have poles and zeros at the same location (see Example 5.10). This is impossible in the scalar case, where H{s) is a 1X1 matrix, since in this case common factors in the numerator and denominator will cancel in the process of forming H{s). In state-space terms, this result can be expressed as follows.
307 CHAPTER 3:
Controllability, Observability, and Special Forms
308 Linear Svstems
LEMMA 5.1. In the SISO case, if A is both a zero of the system and an eigenvalue of "^ ^^^ P^^^ ^^ ^^^ system), then A must be an input- and/or output-decoupling zero (uncontrollable and/or unobservable eigenvalue). This is not necessarily true in the MIMO case. Proof, Assume that A is not an input- and/or output-decoupling zero, or equivalently, that all eigenvalues of A are controllable and observable. Then all eigenvalues of A will be poles of H{s) and all zeros of the system will be zeros of H{s). This is not possible, however, since by definition the numerator and denominator of H{s) cannot have a common factor [presently, {s - A)]. Therefore, A must be an input- and/or output-decoupling zero. • s I s+ s
1
0
the poles and zeros were determined 1 ^ +1 -s + I s^ in Example 5.6 to be {poles of H{s)} = {0, 0, -1} and {zeros of H(s)} = {0, -1}. Note that in this case the poles and zeros are at the same locations, 0 and - 1 . • EXAMPLE 5.10. For H{s)
We conclude by noting that, as will be shown in Section 4.2 of Chapter 4, one can arbitrarily assign values to the controllable eigenvalues, and to a certain extent, one can alter their corresponding eigenvectors, using linear state feedback. In the scalar case, assigning a closed-loop eigenvalue at an open-loop zero location guarantees that the eigenvalue will become unobservable (linear state feedback does not alter controllability) and will not appear in the transfer function (refer to Exercises 4.6 and 4.7 in Chapter 4). In the MIMO case, the eigenvectors of the closed-loop eigenvalues must also be assigned appropriately for the eigenvalues to become unobservable and cancel out when forming the transfer function matrix (see Exercise 4.19 in Chapter 4). 3.6 SUMMARY In this chapter the system properties of reachability (or controllability-from-theorigin) and controllability (-to-the-origin), together with the dual properties of observability and constructibility, respectively, were developed. These concepts were introduced using discrete-time time-invariant systems (Subsection 3.1 A) and were further developed in Part 1 of the chapter (Sections 3.2 and 3.3). In Section 3.2, the reachability Gramian of a continuous-time system was used to derive inputs that transfer the state of the system from one desirable vector value to another. This was accomplished for both time-varying and time-invariant systems. The time-invariant case was developed in Subsection 3.2B, so that it may be studied independently. It was shown that for continuous-time systems, reachability implies controllability, and vice-versa; however, for discrete-time systems, although reachability always implies controllability, controllability may not imply reachability (unless A has full rank). Analogous results were developed in Section 3.3 regarding observability and constructibility. In Part 2, Section 3.4, useful special forms for (continuous-time and discretetime) state-space descriptions of time-invariant systems were developed. The standard forms for uncontrollable (resp., unobservable) systems lead to better understanding of the relationships between state-space and transfer function descriptions of systems. The controller and observer forms provide important structural
information about a system that is useful in state-space realizations of transfer function matrices and in feedback control. Polynomial matrix fractional descriptions of transfer matrices were also introduced, by the structure theorem. Finally, poles and zeros of systems were addressed, in Section 3.5. They were introduced using the Smith form of polynomial matrices and the Smith-McMillan form of transfer function matrices. The reachability (controllability) and observability (constructibility) Gramians played an important role in this chapter. These are now summarized, for convenience. Summary of Gramians Introduced in This Chapter Reachability Gramians rh A.
Wr(to,tl)=
(2.11)
^{ti,T)B{T)B^{T)^{ti,T)dT. J to
B.
WM T) = f e^^-'^^BB^e^^~'^^' Jo
C.
Wr{0,K) = ^A^-^'^^^BB^(A^f-^'+^^
(2.29)
dr.
K-l
K-\
=
^A'BB^(A^y.
(2.64)
1 =0
ControUabihty Gramians A.
Wc(toJl)=
^{tQ,T)B{T)B^{T)^^{tQ,T)dT.
(2.19)
Jto rT
B.
WMT)=^
(2.41)
e-'^^BB^e-'^^^dr. Jo K-l
C.
WM K) = ^ A-~^'^^^BB^{A^)-^'^^\ /=o Observabihty Gramians A.
W,(ro, h)=
\A\ T^ 0.
V ^\T, ro)C^(T)C(T)c&(T, ro) d7.
(2.66)
(3.5)
Jto
B.
Wo(0,T)=l
(3.22)
e^^'C^Ce^'dr. Jo K-\
C.
Wo{0,K) =
(3.49)
^(A^yC^CA\ i=0
Constructibility Gramians A.
Wcnito, tl)=
\' 0^(T,
^I)C^(T)C(T)0(T,
ti)dT.
(3.11)
Jto
B.
Wcn(0,T)=
l' Jo
C.
Wcn(0, K) = ^(A^)"^^'+^^C^CA-^^+^\ i=o
(3.32)
e^'^'-^^C^Ce^^'-^Ur. \A\ 7^ 0.
(3.54)
309 CHAPTER 3:
Controllability, Observability, and Special Forms
310 Linear Systems
where the Gramians in A, B, and C in each case are the Gramians of the systems A. B.
i: = A(t)x + B(t)u, y = Cx(t) + D(t)u. X = Ax -\- Bu, y = Cx -\- Du.
C.
x(k + I) = Ax(k) + Bu(k\ y(k) = Cx(k) + Du(k), respectively.
In B and C the controllability and observability matrices are, respectively, ^ - [B,AB,.,.,A''-^Bl
€ = [ C ^ , ( C A / , ...,(CA^~i)^]^.
Note that reachability is the dual concept to observability and controllability is dual to (re)constructibility. 3.7 NOTES The concept of controllability was first encountered as a technical condition in certain optimal control problems and also in the so-called finite-settling-time design problem for discrete-time systems (see Kalman [6]). In the latter, an input must be found that returns the state XQ to the origin as quickly as possible. Manipulating the input to assign particular values to the initial state in (analog-computer) simulations was not an issue since the individual capacitors could initially be charged independently. Also, observability was not an issue in simulations due to the particular system structures that were used (corresponding, e.g., to observer forms). The current definitions for controllability and observability and the recognition of the duality between them were worked out by Kalman in 1959-1960 (see Kalman [9] for historical comments) and were presented by Kalman in [7]. The significance of realizations that were both controllable and observable (see Chapter 5) was established later in Gilbert [3], Kalman [8], and Popov [12]. For further information regarding these historical issues, consult Kailath [5] and the original sources. Note that [5] has extensive references up to the late seventies with emphasis on the time-invariant case and a rather complete set of original references together with historical remarks for the period where the foundations of the state-space system theory were set, in the late fifties and sixties. Special state-space forms for controllable and observable systems obtained by similarity transformations are discussed at length in Kailath [5] (refer also to the discussion on various canonical forms). Wolovich [19] discusses the algorithms for controller and observer forms and introduces the Structure Theorems. The controller form is based on results by Luenberger [11] (see also Popov [13]). A detailed derivation of the controller form can also be found in Rugh [16]. Original sources for the Canonical Structure Theorem include Kalman [8] and Gilbert [3]. The eigenvector and rank tests for controllability and observability are called PBH tests in Kailath [5]. Original sources for these include Popov [14], Belevich [1], and Hautus [4]. Consult also Rosenbrock [15], and for the case when A can be diagonalized via a similarity transformation, see Gilbert [3]. Note that in the eigenvalue/eigenvector tests presented herein the uncontrollable (unobservable) eigenvalues are also explicitly identified, which represents a modification of the above original results.
The Brunovsky canonical form is developed in Brunovsky [2]. The fact that the controllability indices appear in the work of Kronecker was recognized by Rosenbrock [15] and Kalman [10]. For an extensive introductory discussion and a formal definition of canonical forms, see Kailath [5]. Note that certain special forms exist for time-varying systems as well, but they are not considered here. Multivariable zeros have an interesting history. For a review, see Schrader and Sain [17] and the references therein. Refer also to Vardulakis [18]. 3.8 REFERENCES 1. V. Belevich, Classical Network Theory, Holden-Day, San Francisco, 1968. 2. P. Brunovsky, "A Classification of Linear Controllable Systems," Kybernetika, Vol. 3, pp. 173-187, 1970. 3. E. Gilbert, "Controllability and Observability in Multivariable Control Systems," SIAM J. Control Vol. 1, pp. 128-151, 1963. 4. M. L. J. Hautus, "Controllability and Observability Conditions of Linear Automonous Systems," Proc. Koninklijke Akademie van Wetenschappen, Serie A, Vol. 72, pp. 4 4 3 448, 1969. 5. T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980. 6. R. E. Kalman, "Optimal Nonlinear Control of Saturating Systems by Intermittent Control," IRE WESCON Rec, Sec. IV, pp. 130-135, 1957. 7. R. E. Kalman, "On the General Theory of Control Systems," in Proc. of the First Intern. Congress on Automatic Control, pp. 481-493, Butterworth, London, 1960. 8. R. E. Kalman, "Mathematical Descriptions of Linear Systems," SIAM J. Control, Vol. 1, pp. 152-192, 1963. 9. R. E. Kalman, Lectures on Controllability and Observability, C.I.M.E., Bologna, 1968. 10. R. E. Kalman, "Kronecker Invariants and Feedback," in Ordinary Differential Equations, L. Weiss, ed., pp. 459-471, Academic Press, New York, 1972. 11. D. G. Luenberger, "Canonical Forms for Linear Multivariable Systems," IEEE Transactions on Automatic Control, Vol. 12, pp. 290-293, 1967. 12. V. M. Popov, "On a New Problem of StabiUty for Control Systems," Autom. Remote Control, pp. 1-23, Vol. 24, No. 1, 1963. 13. V. M. Popov, "Invariant Description of Linear, Time-Invariant Controllable Systems," SIAM Journal of Control and Optimization, Vol. 10, No. 2, pp. 252-264, 1972. 14. V M. Popov, Hyperstability of Control Systems, Springer-Verlag, Berlin, 1973. 15. H. H. Rosenbrock, State-Space and Multivariable Theory, Wiley, New York, 1970. 16. W. J. Rugh, Linear System Theory, Prentice-Hall, Englewood Chffs, NJ, 1993. 17. C. B. Schrader and M. K. Sain, "Research on System Zeros: a Survey," Int. Journal of Control, Vol. 50, No. 4, pp. 1407-1433, 1989. 18. A. I. G. Vardulakis, Linear Multivariable Control. Algebraic Analysis and Synthesis Methods, Wiley New York, 1991. 19. W. A. Wolovich, Linear Multivariable Systems, Springer-Verlag, New York, 1974.
3.9 EXERCISES 3.1. (a) Let %k = [B, AB,...,
A^'^B], where
^ ( ^ ^ ) = ^{%n) for k>n,
A^R"><«,igG/?"^'".Showthat and
k) C ^{%n) for
k
311 CHAPTER 3: Controllability, Observability, and Special Forms
312
A srT (b) Let 0^ = [C^, (CAf,...,
Linear Systems
(CA'^-^ff,
M{^k) = >r(0„) for k> n,
where A E R^''^ C E /?^^^ Show that and
J{(€k) D M'(€n) for k
3.2. Consider the state equation x = Ax + Bu, where 0 A
3w2
0 0
1 0 0 0 0 0 —2w 0
0 o' 1 0 B = 0 0 0 1
0 2w 1 0
which was obtained by linearizing the nonlinear equations of motion of an orbiting satellite about a steady-state solution. In the state x = [xi, X2, ^3, ^4]^, xi is the differential radius, while X3 is the differential angle. In the input vector u = [i/i, U2V, u\ is the radial thrust and 1/2 is the tangential thrust. (a) Is this system controllable from ullf
y = yi
is the system observable
from yl (b) Can the system be controlled if the radial thruster fails? What if the tangential thruster fails? (c) Is the system observable from yi only? From y2 only? 3.3. Consider the state equation (a) lfx(0) =
i
0
0
-1
derive an input that will drive the state to
in T sec.
5 , plot u(t), xi(t), X2(t) for T = 1, 2, and 5 sec. Comment on the -5 magnitude of the input in your results.
(b) For x{0)
"1 1 0" 0' 3.4. Consider the state equation x(/:+l) = 0 1 0 x{k) + 1 u(k),y(k) .0 0 1_ .1. (a) Is x^
=
1 0
1 0 x{k). 1 0
reachable? If yes, what is the minimum number of steps required to
transfer the state from the zero state to x^ ? What inputs do you need? (b) Determine all states that are reachable. (c) Determine all states that are unobservable. (d) If i: = Ax + Bu is given with A, B as in (a), what is the minimum time required to transfer the state from the zero state to x^l What is an appropriate u(t)? 3.5. Output reachability (controllability) can be defined in a manner analogous to state reachability (controllability). In particular, a system will be called output reachable if there exists an input that transfers the output from some yo to any ^^i in finite time. Consider now a discrete-time time-invariant system x{k + 1) = Ax(k) + Bu(k), y(k) = Cx(k) + Du(k) with A E /?">'^ B E /?"^^, C E 7?^^", and D E RP'''^. Recall that k-i
y(k) = CA^x(O) + ^
CA^-^'-'^^Buii) + Du(k).
(a) Show that the system {A, B, C, D} is output reachable if and only if rank [D, CB, CAB,...,
CA'^'^B] = p.
Note that this rank condition is also the condition for output reachabihty for continuous-time time-invariant systems x=Ax + Bu,y = Cx + Du. It should be noted that, in general, state reachabihty is neither necessary nor sufficient for output reachabihty. Notice for example that if rank D = p then the system is output reachable, (b) Let D = 0. Show that if (A,5) is (state) reachable, then {A,B,C,D} is output reachable if and only if rank C = p. n 0 0] 11 (c) L e t A = 0 - 2 0 , 5 = 0 C = [1,1,0], and D = 0.
[o (i) (ii)
0 -ij
[i^
Is the system output reachable? Is it state reachable? Let x(0) = 0. Determine an appropriate input sequence to transfer the output to yi = 3 in minimum time. Repeat for x(0) = [1,-1,2] .
3.6. (a) Given x = Ax-^Bu,y = Cx^Du, show that this system is output reachable if and only if the rows of the /? x m transfer matrix H{s) are linearly independent over the field of complex numbers. In view of this result, is the system 1 1 5 + 2 output reachable? His) s (b) Similarly, for discrete-time systems, the system is output reachable if and only if the rows of the transfer function matrix H{z) are hnearly independent over the field of complex numbers. Consider now the system of Exercise 3.5 and determine if it is output reachable. 3.7. Show that the circuit depicted in Fig. 3.6 with input u and output y is neither state reachable nor observable but is output reachable.
"O
T y
1
FIGURE 3.6 Circuit for Exercise 3.7
3.8. A system x = Ax-^Bu,y = Cx + Du is called output function controllable if there exists an input u{t),t G [0,oo), that will cause the output y{t) to follow a prescribed trajectory for 0 < ^ < ^o, assuming that the system is at rest at ^ = 0. It is easiest to derive a test for output function controUabihty in terms of the /? x m transfer function matrix H{s), and this is the approach taken in the following. We say that the m x /? rational matrix HR(S) is a right inverse ofH{s) if H{s)HR{s)=Ip. (a) Show that the right inverse HR{S) exists if and only if rank H{s) = p. Hint: In the sufficiency proof, select///? = H^{HH^)-\ the (right) pseudoinverse of//. (b) Show that the system is output function controllable if and only if H{s) has a right inverse HR{S). Hint: Consider y = Hu. In the necessity proof, show that if rank H
313 CHAPTER 3 :
Controllability, Observability, and Special Forms
314 Linear Systems
(c) Show that the left inverse Hi{s) of H{s) exists if and only if rankH(s) = m. Hint: This is the dual result to part (a). \s + I l l (d) Let H(s) = and characterize all inputs u(t) that will cause the system s s_ 1/^. (at rest at r = 0) to exactly follow a step, _y(^) Part (d) points to a variety of questions that may arise when inverses are considered, including: Is HR(S) proper? Is it unique? Is it stable? What is the minimum degree possible? 3.9. Consider the system x = Ax-\- Bu, y = Cx. Show that output function controllability implies output controllability (-from-the-origin, or reachability). 3.10. Given 4 ^ + 1 ) =
1 0
1 x(k) + 1
u(k), y(k) ~
conditions. (a) Is there a sequence of inputs {w(0), u(l),...}
0'
W
to
0'
w
x(k), and assume zero initial
that transfers the output from ^(0)
in finite time? If the answer is yes, determine such a sequence.
(b) Characterize all outputs that can be reached from the zero output iy{0) =
m
one step. 0 0 1 x(t) + u(t). 0 1 (a) Show that it is controllable at any ^o E (-00,00).
3.11. Consider the state equation x(t) =
(b) Suppose we are interested only in X2(t) \x(t) =
xiit)
Consider, therefore.
xiit) = X2{t) + e~^u{t). Is it possible to determine u{t) so that the state X2it) is transferred from X20 at ^ = ^ [•^2(^0) = -^20] to the zero state at some t = t\ [^2(^1) = 0] and then stay there? If the answer is yes, find such a u{t). (c) In (b), let ^ = 0 and study the effects of the sizes of t\ and XQ on the magnitude of u{t). (d) For the system in (b), determine, if possible, a u{t) so that the state is transferred from xo at ^ = ^0 to x^ at ^ = t\ and then stay there. 3.12. Let F{t) E C"^'^ be a matrix with fi{t) in its /th row. Let f^ G C{R, ^ ^ ) . It was shown that the set fi{t), i = 1 , . . . , n, is linearly independent on [ti, ^2] over the field of complex numbers if and only if the Gram matrix W(ti, ^2) is nonsingular (see Lemma 2.7). If the fi(t), i = 1 , . . . , n, have continuous derivatives up to order (n - 1), then it can be shown that they are linearly independent if for some ^0 ^ [h, h]^ rank[F(to),F^'\to),...,F^''-'\to)]
= n.
If fi(t), i = \, ...,n, are analytic on [t\, ^2], then they are linearly independent if and only if for any fixed to E [^1, ^2], rank [Fit^), F^'\tol...,
F^^'-'Xtol...]
= n.
(a) The above results can be used to derive alternative tests for controllability. In particular, consider the state equation X = A(t)x + B(t)u,
where A(t) E /?"^" and B(t) E /?"^'^ have continuous derivatives up to order (n - 1). Then it can be shown that (A(t), B{t)) is controllable at time ^ if there exists finite ^i > ^ such that rank[Mo{t\),M\{ti),
...,Mn-\{t\)]
= n,
where the Mk{t) G /?"^" are defined by M,+i(0 =
-A{t)Mk{t)+j^Mk{t),
0, 1, . . ., /2 - 1,
with Mo(0 = B{t). (i) Prove this result. (ii) Show that the system i: = A{t)x + B{i)uW\i\\A{t)
't = 0 0
1 0" "0" t 0 .B{t) = 1 0 ^2 .1.
is controllable at any tQ. (b) The results for linear independence of vectors fi(t),i = 1 , . . . , n, can also be used to derive conditions for certain specialized types of controllability. In particular, given X = A(t)x + B(t)u as in (a), a system is called differentially controllable at to, when the transfer from any x(to) = xo to xi can be accomplished in an arbitrarily small interval of time. Note that this may lead to large input magnitudes. It can be shown that when A(0, B{t) are analytic on (-oo, ^), the system is differentially controllable at every t E (-00,00), if and only if for any fixed ^0 G ( - 0 0 , 00),
rank [Mo(to), Mi(to),...,
Mn-i(to), ...] = «.
The system is instantaneously controllable, if and only if for allt G (-00, 00), rank [Mo(t),..., M„_i(0] = n. In this case the transfer of the states can be achieved instantaneously at any time by using inputs that include 6-functions and their derivatives up to order of n - 1. Note that in general instantaneous controllability implies differential controllability, which in turn, implies controllability. In the case when A(t), Bit) are analytic on (-00, 00), as above, then if {A{t), B(t)) is controllable at some point, it is differentially controllable at every t G (-00, 00). Show that in the time-invariant case x = Ax + Bu, controllability of (A, B) always implies both differential and instantaneous controllability; i.e., if a state transfer is possible at all, it can be achieved in an arbitrarily small time interval or even instantaneously if the 5-function and its derivatives are used. For the latter case, see T. Kailath, Linear Systems, Prentice-Hall, 1980, for further details. Remark: In controllability, the transfer of the state occurs in finite time, but the time interval may be very large. In differential controllability, the transfer of the state is possible in arbitrarily small intervals of time; however, this may lead to very large input magnitudes (see Example 2.1). When the system x = A{t)x + B(t)u is uniformly controllable, the transfer of the states can be achieved in some finite time interval using an input with magnitude not arbitrarily large. Note that uniform controllability implies controllability. Uniform controllability is useful in optimal control theory. For additional discussion on differential and uniform controllability, see C. T. Chen, Linear System Theory and Design, Holt, Rinehart and Winston, 1984, and the references cited therein. Note that dual results also exist for differential, instantaneous, and uniform observability.
315 CHAPTERS:
Controllability, Observability, and Special Forms
316 Linear Systems
1 0 0
3.13. Suppose that for the system x{k + 1) known that y{^) = y{l) = y{2) =
1 0 1 0 x{k\ 0 1
y(k)
x(k) it is
. Based on this information, what can be said
about the initial condition x(0)? 3.14. (a) Consider the system x = Ax + Bu, y = Cx + Du, where (A, C) is assumed to be observable. Express x(t) as a function of y(tX u(t) and their derivatives. Hint: Write y(tX y^^\tl..., /^-i>(0 in terms of x(t) and u(t), u^^\t),..., u^^-^\t) (x(t) G /?"). (b) Given the system x = Ax + Bu, y = Cx + Du with (A, C) observable. Determine x(0) in terms of y(t), u{t) and their derivatives up to order n - \. Note that in general this is not a practical way of determining x(0), since this method requires differentiation of signals, which is very susceptible to measurement noise. (c) Consider the system x{k + 1) = Ax{k) + Bu(k), y(k) = Cx(k) + Du(k), where (A, C) is observable. Express x(k) as a function of y(k), y(k + I),..., y(k + n - 1) and u{k), u(k+ 1 ) , . . . , y(k + n- 1). Hint: Express y(k),..., y(k + n- 1) in terms of x(k) and u(k), u(k-\-1),..., u(k+n-1) [x(k) G i?"]. Note the relation to expression (3.48) in Section 3.3. 3.15. Write software programs to implement the algorithms of Section 3.4. In particular: (a) Given the pair (A, B), where A G /?"><", B G 7?"^'" with rank [B, AB,...,
A""^5] = nr < n,
reduce this pair to the standard uncontrollable form A = PAP-^ =
Ai
0
An
B = PB
where (Ai, Bi) is controllable and Ai G /?«'^x«^ Bi G R^rXm^ (b) Given the controllable pair (A, B), where A G /?"><", 5 G i?^^'^ with ran^ 5 = m, reduce this pair to the controller form Ac = PAP'^, Be = PB. 3.16. Determine the uncontrollable modes of each pair (A, B) given below by (a) reducing (A, 5), using a similarity transformation, (b) using eigenvalue/eigenvector criteria.
1 0 0" "1 0" A = 0 - 1 0 ,B = 0 1 0 2. 0 .0 0_
and
0 0 A = 0 0
0 0 0 0
1 1 0 0
o"
0 r
0 0 0 ,B = 1 0 0 -1 0 0
3.17. Consider the system x = Ax + Bu, y = Cx + Du. (a) Show that only controllable modes appear in e^^B, and therefore in the zero-state response of the state. (b) Show that only observable modes appear in Ce^\ and therefore, in the zero-input response of the system. (c) Show that only modes that are both controllable and observable appear in Ce^^B, and therefore, in the impulse response and the transfer function matrix of the system. Consider next the system x{k + 1) = Ax{k) + Bu(k), y(k) - Cx(k) + Du(k). (d) Show that only controllable modes appear in A^B, only observable modes in CA^, and only modes thrt are both controllable and observable appear in CA^B [that is, inHiz)].
0 0" 2 0 ,B = 0 -1. obtained in (d). Compare these
(e) Let A =
1" 0 C - [1, 1, 0], and D - 0. Verify the results .1. with Example 4.10.
3.18. Reduce the pair 0 0 3 0 1 1 1 0
A =
1 -3 4 -1
0 0 1 0 B = 0 1 0 0
0 1 -1 0
into controller form Ac = PAP \ Be = PB. What is the similarity transformation matrix in this case? What are the controllability indices? 3.19. Let A = Ac + BcAmmdB = 5^5^, where the A ^ ^ c are as in (4.63) with A^ G T?'"^^", Bm G /?^x^, and \Bm\ ^ 0. Show that (A, B) is reachable with controllabihty indices juii. Hint: Use the eigenvalue test to show that (A, B) is reachable. Use state feedback to simplify (A, B) (see Exercise 3.21) and show that the /Xi are the controllability indices. 3.20. Show that the controllability indices of the state equation x = Ax-\-BGv, where |G| T^ 0 and (A, B) is reachable, with A G /?«x", B G R^^^^^ are the same as the controllability indices of x = Ax + Bu, within reordering. Hint: Write %k = [BG, ABC,..., A^-^BG] = [5, A 5 , . . . , A^~^B] • [block diag G]=%k' [block diag G] and show that the number of linearly dependent columns in A^5G that occur while searching from left to right in %n is the same as the corresponding number in %n3.21. Consider the state equation x = Ax + Bu, where A G /?">^", B G R"^"^ with (A, B) reachable. Let the linear state-feedback control law be w = F ^ + Gv, F G /^'^x^, G G f^mxm ^'^^^ 1^1 ^ Q 3how that (a) (A + BF, BG) is reachable. (b) The controllability indices of (A + BF, B) are identical to those of (A, B). (c) The controllability indices of (A + BF, BG) are equal to the controllability indices of (A, B) within reordering. Hint: Use the eigenvalue test to show (a). To show (b), use the controller forms in Section 3.4. 3.22. Show that if (A, B) is controllable (-from-the-origin), where A G /?"^^ and B G i?"^^, and rank B = m, then rank A > n — m. 3.23. Consider
0
0
Br =
1 -OLn~l}
0 1
Show that
[5„A,5„...,Ar'5c]
0 0
0 0
0 0
1 c\
0
0
1
•••
c„-3
0
1
Ci
•••
Cn-2
1
Ci
C2
•••
Cn-l
317 CHAPTER 3 :
Controllability, Observability, and Special Forms
318
where Q = -Zf=oQ;„-,-
, 7 1 - 1 , with Co = 1. Also, show that
Linear Systems
•• Oil
^
OL3
an-i 1
1 01
-1 O^n-l
1
1
0
3.24. Given A G Z^^^^, and B G /e^^^'", letra/zy^% = n, where ^ = [5, AB,..., A^^i^]. Consider i G /?^>'^ B G i?'^^^^ with rank % = n, where ^ = [B,AB,..., A'^'^B], and assume that P G Z?"^'' with det P 7^ 0 exists such that
Show that ^ = P 5 and A = PAP'K Hint: Show that (PA - AP)^ = 0. 3.25. Show that the matrices Ac == PAP~\ Be = P 5 are (a) given by (4.47) if P is given by (4.48), (b) given by (4.50) if 2 ( = P'^) is given by (4.49), (c) given by (4.52) if e ( = p-^) is given by (4.51). 3.26. In the circuit of Example 4.9, let R1R2C = L and Ri = R2 = R. Determine x(t) = [xi{t), X2(t)Y and i{t) for unit step input voltage, v(0, and initial conditions x(0) = [a, b]^. Comment on your results. 3.27. Consider the pair (A, b), where A G T?"^'^, b G R'\ Show that if more than one linearly independent eigenvector can be associated with a single eigenvalue, then (A, b) is uncontrollable. Hint: Use the eigenvector test. Let vi, V2 be linearly independent left eigenvectors associated with eigenvalue Ai = A2 = A. Notice that if vib = a i and vib = 0L2, then {a7^v\ - a^^V2)b = 0.
ro
-1
1
3.28. (a) Consider the state equation x = Ax + Bu, x(0) = XQ, where A =
ri 01 1 1 . Determine x{t) as a function of u(t) and XQ, and verify that the .1 2. uncontrollable modes do not appear in the zero-state response, but do appear in the zero-input response (see Example 4.1). (b) Consider the state equation x(k + 1) = Ax(k) + Bu(k) and ^(0) = XQ, where A and B are as in (a). Demonstrate for this case results corresponding to (a). In (a) and (b), determine x(t) and x(k) for unit step inputs and x(0) = [1,1, 1]^. andB =
3.29. (a) Consider the system i = Ax+Bu,y B =
= Cxwithx(O) = xo, where A
•2
-3J
and C = [1, 1]. Determine y(t) as a function of u(t) and XQ, and verify
that the unobservable modes do not appear in the output (see Example 4.3). (b) Consider the system x(k + 1) = Ax(k) + Bu(k), y(k) = Cx(k) with x(0) = XQ, where A, B, and C are as in (a). Demonstrate for this case results which correspond to (a). In (a) and (b), determine and plot y(t) and y(k) for unit step inputs and x(0) = 0. 3.30. Consider the system x(k + 1) == Ax(k) + Bu(k), y(k) = Cx(k), where 1 0 0 A = 0 0
1 2
0
0
_ 1 1
B
C = [1,1,0].
Determine the eigenvalues that are uncontrollable and/or unobservable. Determine x(k), y(k) for /: > 0, given x(0) and u(k), k > 0, and show that only controllable eigenvalues (resp., modes) appear in A^B, only observable ones appear in CA^, and only eigenvalues (resp., modes) that are both controllable and observable appear in CA^B[mH{z)l 3.31. For the system x = Ax + Bu, y = Cx, consider the corresponding sampled-data system x(k + 1) = Ax(k) + Bu(k), y(k) = Cx(k), where A = e^
B =
'dr
B,
and
c =a
(a) Let the continuous-time system {A, B, C] be controllable (observable) and assume it is a SISO system. Show that {A, B, C} is controllable (observable) if and only if the sampling period T is such that 2'Trk
Im (Xi - Xj) 7^ ~^^'
where /: = ± 1 , ± 2 , . . . whenever 7?^ (A/ - Xj)
0,
where {AJ are the eigenvalues of A. Hint: Use the PBH test. Also, consult Appendix D of C. T. Chen, Linear System Theory and Design. Holt, Rinehart, and Winston, 1984. (b) Apply the results of (a) to the double integrator—Example 7.6 in Chapter 2— 0 I 0 0 1 C = [1,0], and also to A where A , B = , B = 0 0 [-1 oj C = [1, 0]. Determine the values of T that preserve controllability (observability).
U\
"1 0~ 1 0 ri 1 0] 0 -1 x + 0 1 ^, y = 1 0 0 0 0 LO OJ (a) Determine the uncontrollable and the unobservable eigenvalues (if any). (b) What is the impulse response of this system? What is its transfer function matrix? (c) Is the system asymptotically stable?
3.32. Given is the system x
3.33. Consider the system x(k + 1) =
x(k) +
that it is known that for zero input, y(0) If yes, find x(0) and verify your answer.
u(k), y(k) = [I, 3]x(k). Suppose
1 and _y(l) = 1. Can x(0) be determined?
0 3.34. Given is the transfer function matrix H(s)
s+1
s+2
0 (a) Determine the Smith-McMillan form of H(s) and its characteristic (pole) polynomial and minimal polynomial. What are the poles of H(s)7 (b) Determine the zero polynomial of H(s). What are the zeros of H(s)l c2
3.35. Let H(s) =
1
s^ s+ 1
(a) Determine the Smith-McMillan form of H(s) and its characteristic (pole) polynomial and minimal polynomial. What are the poles of H(s)7 (b) Determine the zero polynomial of H(s). What are the zeros of H(s)l 3.36. A rational function matrix R(s) may have, in addition to finite poles and zeros, poles and zeros at infinity (s = oo). To study the poles and zeros at infinity, the bilinear
319 CHAPTERS:
Controllability, Observability, and Special Forms
320
transformation biw +
Linear Systems
s =
aiw + ao with ai 7^ 0, biao - h^ax y^ 0 may be used, where b\lai is not a finite pole or zero of R{s). This transformation maps the point s = b\la\ iow = ^ and the point of interest, 5- = 00, to w = -aja\. The rational matrix R{\v) is now obtained as R{w) = R
1\W •
aiw -\- ao
and the finite poles and zeros of R(w) are determined. The poles and zeros at w = -ao/a\ are the poles and zeros of R(s) at 5* = oo. Note that frequently a good choice for the bilinear transformation is ^ = 1/w, that is, bi = 0, bo = 1 and ai = I, ao = 0. (a) Determine the poles and zeros at infinity of 1 1 Ri(s) = Rsis) = Riis) s+ 1 s+ r Note that a rational matrix may have both poles and zeros at infinity, (b) Show that if R(s) has a pole at 5" = oo, then it is not proper (lim^^oo R(s) •
-).
3.37. Determine the poles and zeros at infinity of the transfer functions in Examples 5.1,5.2, 5.6, and 5.7. 3.38. (Spring mass system) Consider the spring mass given in Exercise 2.69 in Chapter 2. (a) Is the system controllable from [/i, /i]^? If yes, reduce (A, B) to controller form. (b) Is the system controllable from input fi only? Is it controllable from /2 only? Discuss your answers. ri 0 0 01 (c) Let y = Cx with C = Is the system observable from j ? If yes. 0 1 0 0 reduce (A, C) to observer form. 3.39. (Aircraft dynamics) Consider the state-space description of the lateral motion of an aircraft in Exercise 2.76 in Chapter 2. (a) Is the system controllable from [8A, 8R]^7 If yes, reduce (A, B) to controller form. (b) Is the system controllable using only the ailerons? Is the system controllable using only the rudder? Discuss your answers. ro 1 0 01 (c) Let y = Cx with C = Is the system observable from yl If yes.
[o 0 1 oj
reduce (A, C) to observer form.
CHAPTER 4 =1 ^lir?:S-«^|t; • > # l ^ i -^ *< - >V«ct<-<<;^'^-'^"^-;;¥
State Feedback and State Observers
Feedback is a fundamental mechanism arising in nature and is present in many natural processes. Feedback is also common in manufactured systems and is essential in automatic control of dynamic processes with uncertainties in their model descriptions and their interactions with the environment. When feedback is used, the actual values of system variables are sensed, fed back, and used to control the system. Hence, a control law decision process is based not only on predictions about the system behavior derived from a process model (as in open-loop control), but also on information about the actual behavior (closed-loop feedback control). A common example of an automatic feedback control system is the cruise control system in an automobile, which maintains the speed of the automobile at a certain desired value within acceptable tolerances. In this chapter feedback is introduced, and the problem of pole or eigenvalue assignment by means of state feedback is discussed at length in Section 4.2. It is possible to arbitrarily assign all closed-loop eigenvalues by linear static state feedback if and only if the system is completely controllable. This relation to controllability is, in fact, the motivation for introducing state feedback at this point. Feedback control is considered again in Chapter 7, where polynomial matrix descriptions are introduced. In the study of state feedback it is assumed that it is possible to measure the values of the states using appropriate sensors. Frequently, however, it may be either impossible or impractical to obtain measurements for all states. It is therefore desirable to be able to estimate the states from measurements of input and output variables that are typically available. In addition to feedback control problems, there are many other problems where knowledge of the state vector is desirable since such knowledge contains useful information about the system. This is the case, for example, in navigation systems. State observers that asymptotically estimate the states from input and output measurements over time are also studied in this chapter. State estimation is related to observability in an analogous way that state feedback control is related to controllability. The duality between controllability and
321
322
Linear Systems
observability makes it possible to easily solve the estimation problem once the control problem has been solved, and vice versa. In this chapter, full-order and reduced-order asymptotic estimators, also called observers, are discussed at length in Section 4.3. Finally, state feedback static controllers and state dynamic observers are combined to form dynamic output feedback controllers. Such controllers are studied in Section 4.4, using both state-space and transfer function matrix descriptions.
4.1 INTRODUCTION A. A Brief Introduction to State Feedback Controllers and State Observers In the following discussion, state feedback and state estimation are introduced for continuous- and discrete-time time-varying and time-invariant systems. We consider systems described by equations of the form i: = A{t)x + B{t)u,
y = C(t)x + D(t)u,
(1.1)
where t G (a, b), some real open interval, and A(t) G R''^'', B{t) G W'^'^^ C(t) G j^pxn^ £)(r) G RP^^^ and u(t) G R"^ are (piecewise) continuous in t on (a, b) (see Chapter 2). Let the input u be determined by a time-varying linear, state feedback control law of the form u = F{t)x + r,
(1.2)
where F{t) G R^^^ is (piecewise) continuous in f G {a, b), the elements of F(t) represent time-dependent gains, and r(t) G R^ is an external input (see Fig. 4.1). Substituting into (1.1), the state-space description of the compensated or closedloop system is given by is given by X = [A(0 + B(t)F(t)]x + B(t)r y - [C(0 + D(t)F(t)]x + D(t)r
(1.3)
We seek to select F(t) so that the closed-loop system has certain desirable qualitative properties. For example, we may wish the closed-loop system to be stable. (For definitions of stability in the time-varying case, refer to Chapter 6.) Stability can be achieved under appropriate assumptions involving certain types of controllability. One way of determining such stabilizing F(t) is to use results from the optimal Linear Quadratic Regulator (LQR) theory, which in fact yields the "best" F(t) in some
System
+T
FIGURE 4.1
sense. Stabilization, or the achievement of other control objectives, for linear timevarying systems via LQR or other methods will not be studied in this chapter. The stabilization of linear time-invariant systems, however, is discussed at great length in Section 4.2. In the time-invariant case we consider systems described by equations of the form X = Ax + Bu,
y = Cx + Du,
where A E /?^x^ B G /^^X"^, C G J?^><^ and D G
RP"""^.
(1.4)
Let
u = Fx + r
(1.5)
represent the linear time-invariant or static state feedback control law. The compensated or closed-loop system is then given by the equations i - (A + BF)x + Br,
y = (C -\- DF)x + Dr
(1.6)
In control design, the gain matrix F is selected so that the closed-loop system (1.6) has desired behavior. In this chapter we are particularly interested in selecting F to stabilize the system, and we will concentrate on the eigenvalue assignment problem. In particular, it is shown in the next section that the controllable eigenvalues of the system [i.e., of the pair (A + BF, B)] can be assigned (or shifted) to arbitrary locations, and therefore, can be assigned to locations that guarantee stable behavior. In view of the results developed in the previous chapter, the central role played by controllability in feedback stabilization is most easily seen in the time-invariant case. Recall that controllability reflects the ability of an input to transfer the state of a system to any desirable (state) value. It was shown in Section 3.4 that given a time-invariant system {A, B, C, D}, there exists a transformation so that the controllable part of the state space can easily be recognized. In other words, there exists a basis for the state space, so that all controllable states have representations of the form [x[, 0^]^ and the uncontrollable state representations are of the form [x[, X2V, where X2 = A2X2, i.e., X2 is independent of u (see Subsection 3.4A). The eigenvalues of A2 are in this case the uncontrollable eigenvalues that correspond to the uncontrollable modes of the system. Since u has no effect on X2, it is reasonable to conjecture that the state feedback control law (which in fact, is a particular choice for u) will not at all affect the eigenvalues of A2, for if it did, it would also affect X2. This conjecture turns out to be true and will formally be shown in the next section. The controllable states can now arbitrarily be shifted, using u. This in turn means that the controllable eigenvalues can be arbitrarily shifted, using w, or more correctly, by appropriately selecting u in xi = AiXi -\- Biu -^ A12X2, xi may be made to behave as if the eigenvalues of Ai were arbitrarily shifted. This is true because if there were restrictions on the location of the eigenvalues, then the state would not be able to be shifted arbitrarily, which is a contradiction. It turns out that linear state feedback control provides such u, and therefore, it can arbitrarily shift all the eigenvalues of a system. This will formally be shown in Section 4.2. A stabilizing state feedback matrix F may also be determined as the solution to an optimal control problem, the Linear Quadratic Regulator (LQR) problem. This is briefly addressed at the end of Section 4.2.
323 CHAPTER 4:
state Feedback and State Observers
324
Linear Systems
We will also consider time-varying discrete-time systems, described by equations of the form x{k + 1) = A{k)x{k) + B{k)u{k),
y(k) = C(k)x(k) + D{k)u{k)
(1.7)
with linear, discrete-time, time-varying state-feedback control law given by u{k) = F(k)x(k) + r(k),
(1.8)
In this case, the closed-loop system is given by x(k + 1) = [A(k) -h B(k)F(k)]x(k)
+ B(k)r(k)
y(k) = [C(k) + D(k)F(k)]x(k)
+ Dik)r(k).
(1.9)
Also, we will consider time-invariant discrete-time systems described by x(k + 1) = Ax(k) + Bu(k),
y(k) = Cx(k) + Du(k)
(1.10)
with linear, discrete-time, time-invariant state-feedback control law (time-invariant) given by u{k) = Fx(k) + r(k\
(1.11)
In this case the closed-loop system assumes the form x(k + 1) - [A + BF]x(k) + Br(k) y(k) = [C + DF]x(k) + Dr(k).
(1.12)
As in the continuous-time case, stabilization is emphasized in this chapter, and corresponding results are developed. In (continuous-time) state-space system representations, if the value of the state at time to, x(to), is known, then the input u(t), t > ^ , uniquely determines x(t) and y(t) for t ^ to. Since knowledge of the initial value of the state is of such importance, methods of determining (estimating) x(to) from input and output measurements have been devised (see Fig. 4.2). Methods of estimating the initial value of the state, x(0), in time-invariant systems are addressed in Section 4.3. In particular, full- and reduced-order state observers are designed that asymptotically estimate the state. Since controllability (-from-the-origin, or reachability) was the key property in state feedback control, the dual property of observabihty, studied in Section 3.3, is the key attribute in state estimation. Recall that observability refers to the ability of determining the state from measurements of the output and input over a finite time interval. (See, for instance. Corollary 3.8 in Chapter 3, where the initial state is determined using the observability Gramian of the system.) Furthermore, since the solution to the optimal u
y Sysiern
State observer
\, *
FIGURE 4.2
LQR control problem leads to optimal state-feedback control law matrices [F(t) or F], a dual optimal estimation problem can be defined, the solution of which provides optimal state-observer gain matrices [K(t) or K]. Corresponding discrete-time results on state estimation are also discussed. In addition, in Section 4.4, state feedback control laws and state estimation are combined to derive a dynamic-observer based controller that receives feedback information, not from the state, but from the outputs of the system to be controlled. These are typically more accessible.
B. Chapter Description In this chapter state feedback controllers and asymptotic state estimators, also called state observers, are introduced and studied. The contents of the chapter were outlined and briefly discussed in the previous subsection (Subsection 4.1 A). A detailed summary follows. A brief introduction to state feedback controllers and state observers for continuous-time and discrete-time systems is given in Subsection 4.1 A. In Section 4.2, state feedback control is studied. Our development focuses on time-invariant systems. First, the need for feedback is demonstrated by a discussion of open- and closed-loop control laws, and their differences when uncertainties are present in the system model and its environment. The stabilization of systems is emphasized. This leads, in the time-invariant case, to the eigenvalue (or pole) assignment problem, which is studied next at length. The flexibility offered in the multi-input case in the choice of the state feedback gains to accomplish control goals, in addition to pole assignment, is stressed. A number of pole assignment methods are introduced, including the eigenvalue/eigenvector assignment method; an additional, historically important method is presented in Exercise 4.2. It is pointed out that stabilization can also be achieved by an optimal control formulation, such as the Linear Quadratic Regulator (LQR), which leads to a stabilizing state feedback control law while attaining additional control goals as well. The LQR problem is briefly discussed for both the continuous- and discrete-time cases. In Section 4.3, full-order, full-state, and partial-state observers for continuousand discrete-time systems are studied. Also, reduced-order and optimal observers are addressed. In the discrete-time case, current state estimators are also introduced. Optimal Linear Quadratic Gaussian (LQG) estimators are briefly discussed. The duality of the state feedback and the state observer problems is emphasized. It is pointed out that the main factor that limits the magnitude of the gains in observers is noise. This is in contrast to the limiting factor in the magnitude of the control gains, which is limitations of the control actuators and of the linear model used to describe the system. In Section 4.4, dynamic state observers are used, together with static state feedback controllers, to derive dynamic output controllers.' The Separation Principle is discussed and the degradation of performance in state feedback control when an observer is used to estimate the state is explained. The analysis is accomplished in both state-space and transfer function frameworks. Furthermore, additional output controller configurations are also derived in terms of state-space representations.
325 CHAPTER 4:
state Feedback and State Observers
326 Linear Systems
C. Guidelines for the Reader In this chapter, Unear state feedback controllers and linear state observers for linear time-invariant systems are studied. Subsection 4,IB describes the contents of the chapter. At a first reading, one can cover selected topics from the material on state feedback and state observers in Sections 4.2 and 4.3, respectively. Regarding Section 4.2 on state feedback, after studying the issues associated with open- and closed-loop control in Subsection 4.2A, one could concentrate on eigenvalue assignment using linear state feedback. In particular, one could study Theorem 2.1 in Subsection 4.2B, which shows that only the controllable eigenvalues can be arbitrarily assigned via state feedback; then one could study the methodologies to assign desired eigenvalues (and in part, the corresponding eigenvectors as well). Regarding Section 4.3 on state observers, at a first reading one could concentrate only on full-order full-state observers for continuous-time systems in Subsection 4.3A. Then one could study dynamic output feedback controllers that are based on state observers in Section 4.4. The degradation of performance when an observer is used to estimate the state in a state feedback controller is also discussed in that section.
4.2 LINEAR STATE FEEDBACK A. Continuous-Time Systems We consider linear, time-invariant, continuous-time systems described by equations of the form x = Ax + Bu,
y = Cx + Du,
(2.1)
where A G /?"><^ B G /e'^x^, C G /^^x^ and D G /?^><^. DEFINITION 2.1. The linear, time-invariant, state feedback control law is defined by u = Fx + r, where F G R^^^ is a gain matrix and r{t) G R^ is an extemal input vector.
(2.2) •
Note that r{t) is an external input, also called a command or reference input (see Fig. 4.1). It is used to provide an input to the compensated closed-loop system and is omitted when such input is not necessary in a given discussion [r{t) = 0]. This is the case, e.g., when the Lyapunov stability of a system is studied. Note that the vector r(t) in (2.2) has the same dimension as u(t). If a different number of inputs is desired, then an input transformation map may be used to accomplish this. The compensated closed-loop system of Fig. 4.1 is described by the equations i = (A + BF)x + Br y = {C + DF)x + Dr,
(2.3)
which were determined by substituting u = Fx ^- r into the description of the uncompensated open-loop system (2.1).
The state feedback gain matrix F affects the closed-loop system behavior. This is accomplished by altering the dynamic effects of the matrices A and C of (2.1). In fact, the main influence of Fis exercised through the matrix A, resulting in the matrix A + BF of the closed-loop system. The matrix F affects the eigenvalues of A + BF, and therefore, the modes of the closed-loop system. The effects of F can also be thought of as restricting the choices for w ( = Fx for r = 0) so that for appropriate F, certain properties, such as asymptotic Lyapunov stability, of the equilibrium x = 0 are obtained. Open- and closed-loop control The linear state feedback control law (2.2) can be expressed in terms of the initial state x(0) = XQ. In particular, working with Laplace transforms, we obtain u = Fx-^f = F[(^/-A)~^xo + (^/-A)~^Bw]-t-f, inviewof ^-x-xo = Ax + Bu, derived from x = Ax + Bu. Collecting terms, we have [/ - F{sl - A)~^B^u = F(sl - A)~^xo + r. This yields u = F[sl - (A + BF)r^xo
+ [/ - F(sl - Ay^BV^r,
(2.4)
where the matrix identities [/ - F(sl - A)-^Br^F(sI - A)"^ = F(sl - Ay^[I BF(sI - A)-i]-i ^ F[sl - (A + BF)]-^ have been used. Expression (2.4) is an open-loop (feedforward) control law, expressed in the Laplace transform domain. It is phrased in terms of the initial conditions x(0) = XQ, and if it is applied to the open-loop system (2.1), it generates exactly the same control action u(t) for r > 0 as the state feedback u = Fx -\- r in (2.2). It can readily be verified that the descriptions of the compensated system are exactly the same when either control expressions, (2.2) or (2.4), are used (verify this). In practice, however, these two control laws hardly behave the same, as explained in the following. First, notice that in the open-loop scheme (2.4) the initial conditions XQ are assumed to be known exactly. It is also assumed that the plant parameters in A and B are known exactly. If there are uncertainties in the data, this control law may fail miserably, even when the differences are small, since it is based on incorrect information without any way of knowing that these data are not valid. In contrast to the above, the feedback law (2.2) does not require knowledge of XQ. Moreover, it receives feedback information from x(t) and adjusts u(t) to reflect the current system parameters, and consequently, is more robust to parameter variations. Of course the feedback control law (2.2) will also fail when the parameter variations are too large. In fact, the area of robust control relates feedback control law designs to bounds on the uncertainties (due to possible changes) and aims to derive the best design possible under the circumstances. The point we wish to emphasize here is that although open- and closed-loop control laws may appear to produce identical effects, typically they do not, the reason being that the mathematical system models used are not sufficiently accurate, by necessity or design. Feedback control and closed-loop control are preferred to accommodate ever-present modeling uncertainties in the plant and the environment. The purpose of controlling a system is to achieve certain control goals. Examples include tracking a given trajectory, regulating the state so that it returns to the origin if it is disturbed, and stabilizing a system. Stabilization is discussed at length in this chapter. Regulation and tracking, together with other control problems, are briefly discussed in Chapter 7, where output feedback compensation, in addition to state feedback compensation, is used.
327 CHAPTER 4 :
State Feedback and State Observers
328 Linear Systems
At this point, a few observations are in order. First, we note that feeding back the ^^ate in synthesizing a control law is a very powerful mechanism, since the state contains all the information about the past history of a system that is needed to uniquely determine the future system behavior, given the input. We observe that the state feedback control law considered presently is linear, resulting in a closed-loop system that is also linear. Nonlinear state feedback control laws are of course also possible. Notice that when a time-invariant system is considered, the state feedback is typically static, unless there is no choice (as in certain optimal control problems), resulting in a closed-loop system that is also time-invariant. These comments justify to a certain extent the choice of linear, time-invariant, state feedback control to compensate linear time-invariant systems. The problem of stabilizing a system by using state feedback is considered next. Stabilization The problem we wish to consider now is to determine a state feedback control law (2.2) having the property that the resulting compensated closed-loop system has an equilibrium x = 0 that is asymptotically stable (in the sense of Lyapunov) when r = 0. (For a discussion of asymptotic stability, refer to Chapters 2 and 6.) In particular, we wish to determine a matrix F G R^^^ so that the system i: = (A + BF)x,
(2.5)
where A G R^^^midB E 7?'^^'^, has equilibrium x = 0 that is asymptotically stable. Note that (2.5) was obtained from (2.3) by letting r = 0. One method of deriving such stabilizing F is by formulating the problem as an optimal control problem, e.g., as the Linear Quadratic Regulator (LQR) problem. This is discussed at the end of this section. We point out that an LQR formulation can also be used to derive stabilizing gains F(t) in the time-varying case so that the equilibrium x = 0 of the system x = [A(t) + B(t)F(t)]x is asymptotically stable. However, this will not be pursued here. Alternatively, in view of Chapter 2 (and Chapter 6), the equilibrium x = 0 of (2.5) is asymptotically stable if and only if the eigenvalues A/ of A + BF satisfy Re A/ < 0, / = 1 , . . . , n. Therefore, the stabilization problem for the time-invariant case reduces to the problem of selecting F in such a manner that the eigenvalues of A + BF are shifted into desired locations. This will be studied in the following subsection. Note that stabilization is only one of the control objectives, although a most important one, that can be achieved by shifting eigenvalues. Since the eigenvalues of a linear system determine its qualitative dynamic behavior (refer to the discussion of modes in Chapter 2), one can attain a number of control goals by shifting of eigenvalues, in addition to stability. Control system design via eigenvalue (pole) assignment is a topic that is addressed in detail in a number of control books.
B. Eigenvalue Assignment Consider again the closed-loop system i = (A + BF)x given in (2.5). We shall show that if (A, B) is fully controllable (-from-the-origin, or reachable), all eigenvalues of A + BF can be arbitrarily assigned by appropriately selecting F. In other words, "the eigenvalues of the original system can arbitrarily be changed in this case." This last
statement, commonly used in the literature, is rather confusing: The eigenvalues of a given system x = Ax + Bu, are not physically changed by the use of feedback. They are the same as they used to be before the introduction of feedback. Instead, the feedback law u = Fx ~\- r, r = 0, generates an input u(t) that, when fed back to the system, makes it behave as if the eigenvalues of the system were at different locations [i.e., the input u(t) makes it behave as a different system, the behavior of which is, we hope, more desirable than the behavior of the original system]. THEOREM2.1. GivenA G /?"><" a n d 5 G Z?"^'", there exists F G /?""><" such that then eigenvalues of A + BF can be assigned to arbitrary, real or complex conjugate, locations if and only if (A, B) is controllable (-from-the-origin, or reachable). Proof. (Necessity) Suppose that the eigenvalues of A + BF have been arbitrarily assigned and assume that (A, B) in (2.1) is not fully controllable. We shall show that this leads to a contradiction. Since (A, B) is not fully controllable, in view of the results in Section 3.4 in Chapter 3, there exists a similarity transformation that will separate the controllable part from the uncontrollable part in (2.5). In particular, there exists a nonsingular matrix Q such that Q-\A
+ BF)Q
= Q-'AQ
+ (Q-'B)(FQ)
Ai + BiFi 0
=
Ai 0
An Ai.
Ai2 + BxF2 A2
{F,,F2\ (2.6)
where [Fi, F2] = FQ and {A\,Bi) is controllable. The eigenvalues of A + BF are the same as the eigenvalues of Q~^{A + BF)Q, which implies that A + BF has certain fixed eigenvalues, the eigenvalues of A2, that cannot be shifted via F. These are the uncontrollable eigenvalues of the system. Therefore, the eigenvalues of A + BF have not been arbitrarily assigned, which is a contradiction. Thus, (A, B) is fully controllable. (Sufficiency) Let (A, B) be fully controllable. Then by using any of the eigenvalue assignment algorithms presented later in this section, all the eigenvalues of A + BF can be arbitrarily assigned. • LEMMA 2.2. The uncontrollable eigenvalues of (A, B) cannot be shifted via state feedback. Proof See the necessity part of the proof of Theorem 2.1. Note that the uncontrollable eigenvalues are the eigenvalues of A2. • EXAMPLE 2.1. Consider the uncontrollable pair (A, 5), where A =
0 -2 , B = 1 -3
. This pair can be transformed to a standard form for uncontrollable systems, namely, -2 1 , from which it can easily be seen that - 1 is the uncontrollable B = 0 -1 eigenvalue, while - 2 is the controllable eigenvalue (see also Example 4.3 in Chapter 3).
A =
NowifF = [f,f2lihendet(sI-(A
+ BF)) = det\ \~ ^\ ]_~^^\ L - 1 - Jl S + 3- f2_ s(-f\ -f2 + 3) + (-f-f2 + 2) = (s+ l)(s + ( - / i - /2 + 2)). Clearly, the uncontrollable eigenvalue - 1 cannot be shifted via state feedback. The controllable eigenvalue - 2 can be shifted arbitrarily to (f + /s - 2) by F = [/i, /2]. •
329 CHAPTER 4 :
State Feedback and State Observers
330 Linear Systems
It is now quite clear that a given system (2.1) can be made asymptotically stable via the state feedback control law (2.2) only when all the uncontrollable eigenvalues of (A, B) are already in the open left part of the s-plane. This is so because state feedback can alter only the controllable eigenvalues. DEFINITION 2.2. The pair (A, B) is called stabilizable if all its uncontrollable eigenvalues are stable. • Before presenting methods to select F for eigenvalue assignment, it is of interest to examine how the linear feedback control law u = Fx + r given in (2.2) affects controllability and observability. We write si -{A + BF) -{C + DF)
B D
si-A
-c
B\
\
D\ [-F
^ ^ I
(2.7)
and note that rank [A/ - (A + BF), B] = rank [XI - A, B] for all complex A (show this). Thus, if (A, B) is controllable, then so is (A + BF, B) for any F. Further, notice that in view of %F = [B, (A + BF)B, (A + BFfB,..., I 0 =
[B,AB,A^B,...,A''~'B]
(A + BF^'^ 'B] ^ FB /
F(A + BF)B FB I
(2.8)
gi(^/r) = ^([B, AB,..., A^-i^]) = S/l(^) (why?). This shows that F does not alter the controllability subspace of the system (see Section 3.2.). This in turn proves the following lemma. LE M M A 2.3. The controllability subspaces of x Ax + Bu and i: = (A + BF)x + Br are the same for any F. Although the controllability of the system is not altered by linear state feedback u == Fx + r, this is not true for the observability property. Note that the observability of the closed-loop system (2.3) depends on the matrices (A + BF) and (C + DF), and it is possible to select F to make certain eigenvalues unobservable from the output. In fact this mechanism is quite common and is used in several control design methods. It is also possible to make observable certain eigenvalues of the open-loop system that were unobservable; for an example see Example 2.8 below. Several methods are now presented to select F so to arbitrarily assign the closedloop eigenvalues. Methods for eigenvalue assignment by state feedback In view of Theorem 2.1, the eigenvalue assignment problem can now be stated as follows. Given a controllable pair (A, B), determine F to assign the n eigenvalues of A + BF to arbitrary real and/or complex conjugate locations. This problem is also known as the pole assignment problem, where by the term "pole" is meant a "pole of the system" (or an eigenvalue of the "A" matrix). This is to be distinguished from the "poles of the transfer function" (see Section 3.5 for the appropriate definitions).
Note that all matrices A, B, and F are real, so the coefficients of the polynomial det [si - (A + BF)} are also real. This imposes the restriction that the complex roots of this polynomial must appear in conjugate pairs. Also, note that if (A, B) is not fully controllable, then (2.6) can be used together with the methods described a little later, to assign all the controllable eigenvalues; the uncontrollable ones will remain fixed (see Theorem 2.1 and Lemma 2.2). It is assumed in the following that B has full column rank, i.e.. rank B = m.
(2.9)
This means that the system x = Ax + Bu has m independent inputs. If rank B = r < m, this would imply that one could achieve the same result by manipulating only r inputs (instead of m > r). To assign eigenvalues in this case, one can proceed by writing A + J5F - A + (BM)(M-^F)
= A + [5i, 0]
A + 5iFi,
(2.10)
where M is chosen so that BM = [Bi, 0] with Bi G T?"^'* and rank Bi = r. Then Fi G R^^^ can be determined to assign the eigenvalues of A + BiFi, using any one of the methods presented next. Note that (A, B) is controllable implies that (A, Bi) is controllable (why?). The state feedback matrix F is given in this case by (2.11)
F =M where F2 G R(^-^)^n is arbitrary.
/ . Direct method. Let F = [fij], i = 1 , . . . , m, j = \,.. .,n, and express the coefficients of the characteristic polynomial of A + BF in terms of fij, i.e., det {si - (A + BF)) = s- + gn-iifiX''
+ • • • + goifijl
Now if the roots of the polynomial (^d(s) = ^" + dn-lS"" ^ + • • • + Ji^ + Jo are the n desired eigenvalues, then the fij, i = \,.. .,m, j = 1, termined so that gk(fij) = dk,
A: = 0, 1, . . . , n - 1.
, n, must be de(2.12)
In general, (2.12) constitutes a nonlinear system of algebraic equations; however, it is linear in the single-input case, m = \. The main difficulty in this method is not so much in deriving a numerical solution for the nonlinear system of equations, but in carrying out the symbolic manipulations needed to determine the coefficients g]^ in terms of the ftj in (2.12). This difficulty usually restricts this method to the simplest cases, with n = 2 or 3 and m = 1 or 2 being typical. EXAMPLE 2.2. For A
[B
we have det (si - A) = s(s - 5/2), and
therefore, the eigenvalues of A are 0 and | . We wish to determine F so that the eigenvalues of A + BF are at - 1 ± 7.
331 CHAPTER 4:
state Feedback and State Observers
332
IfF = [fu f2lihQndet(sI-(A
Linear Systems det
s-\-fi -l-/i
+ BF)) = det
-1
-1 s-2
[fl,f2]] =
-1-/2 = s^ + s{-l - fi - f2) + f\ - Ifi. The desired eigens-2-f2_
values are the roots of the polynomial aM
= (^ - (-1 + Ms - (-1 - j)) ^ s^ + 2s + 2.
Equating coefficients, one obtains -\ — fi - fi = 2, fi - \f2 = 2, a. linear system of equations. Note that it is linear because m = 1. In general one must solve a set of nonlinear algebraic equations. We have
F = [fuf2] =
[-l-'i]
as the appropriate state feedback matrix.
•
2. The use of controller forms. Given that the pair (A, B) is controllable, there exists an equivalence transformation matrix P so that the pair (Ac = PAP~^,Bc = PB) is in controller form (see Section 3.4). The matrices A + BF and P(A + BF)P~^ = PAP~^ + PBFP-^ = Ac -^ BcFc have the same eigenvalues, and the problem is to determine Fc so that Ac + BcFc has desired eigenvalues. This problem is easier to solve than the original one because of the special structures of Ac and Be. Once Fc has been determined, then the original feedback matrix F is given by (2.13)
FrP
We shall now^ assume that (A, B) has already been reduced to (A^, Be) and describe methods of deriving Fc for eigenvalue assignment. Consider first the single-input case (m = 1). We let (2.14)
Fc - [fo> " -y fn-l]' In view^ of Section 3.4, since A^ Be are in controller form, v^e have AcF = Ac + BcFc 0
1
0
0 -ao
0 -ax
1 Oin-\
0 -(<^o - /o)
0 - ( « i - /i)
"o" +
0 1
[/O. •••> fn-\\
0
(2.15) -{pin-\ -
fn-\)
where a/, / = 0 , . . . , n - 1, are the coefficients of the characteristic polynomial of Ac, i.e., det{sl - Ac) = s"" + an-is""'^ + • • • 4- a:i^ + ao-
(2.16)
Notice that ACF is also in companion form and its characteristic polynomial can be written directly as det{sl - ACF) = S- + {dn-i ~ fn-Ds""-' + ' ' ' + (CQ " /Q).
(2.17)
333
If the desired eigenvalues are the roots of the polynomial aais)
= s"" + dn-is""-^
+ • • • + ^o,
(2.18)
then by equating coefficients, fi,i = 0, 1 , . . . , n - 1, must satisfy the relations dt = at - fiJ = 0,1,.. .,n - 1, from which we obtain (2.19)
0,...,n-h
«/ - di.
fi
Alternatively, note that there exists a matrix A^ in companion form, the characteristic polynomial of which is (2.18). An alternative way of deriving (2.19) is then to set AcF = Ac + BcFc = A^, from which we obtain
Fc =
(2.20)
B-'[A.-Aml
where Bm = 1, Aj^ = [-do,..., -d^-i] and Am = [-OCQ, ..., - a „ - i ] . Therefore, Bm, A j ^ , and A ^ are the nth rows of B^ A^, and A^, respectively (see Section 3.4). Relationship (2.20), which is an alternative formula to (2.19), has the advantage that it is in a form that can be generalized to the multi-input case. EXAMPLE2.3. Consider the matrices A
B =
of Example 2.2. Deter-
mine F so that the eigenvalues of A + BF are - 1 ± j , i.e., so that they are the roots of the polynomial ad(s) = s^ + 2s + 2. To reduce (A, B) into the controller form, let ^ - [B, AB] = from which P
1
3
1
3
-2
and^
-1 _
3 -1
[see (4.44) in Chapter 3]. Then P'^ =
qA
-1 i 2
and Ar = PAP-^ =
Br =
1 and Ad^ = [ - 2 , - 2 ] since the -2 characteristic polynomial of A^ is ^^ + 2^ + 2 = ad(s). Applying (2.20), we obtain that
Thus, Am = [0, f ] and B^ = I. Now Aj =
Fc = B-'[Ad^-Am] r
and F = FcP =
[-2,
2J -
3
21 3
1 3
2 3-
2
[-
= [-2,-|] - y ] assigns the eigenvalues of the
closed-loop system at - 1 ± j . This is the same result as the one obtained by the direct method given in Example 2.2, If ad(s) = s^ + dis + do, then Aj^^ = [-do, -d\\ Fc = B-J{Ad^ - Am] = [-^0, -dx - 5/2], and F = FcP =
\[2do-di
-2do - 2di - 5].
In general the larger the difference between the coefficients of ad{s) and Q:(5'), (A^^ -Am), the larger the gains in F. This is as expected, since larger changes require in general larger control action. • Note that (2.20) can also be derived using (4.63) of Section 3.4 in Chapter 3. To see this, write AcF = Ac + BcFc = (Ac + BcAm) + {BcBm)Fc = A , + 5 , ( A ^ +
BmFd
CHAPTER 4 :
State Feedback and State Observers
334 Linear Systems
where A^ Be are defined in (4.63). Selecting Aj = A^ + BcAd^ and requiring ACF = Aj implies Bc[Am + BfnFc] = BcAd^, from which A^ + 5m^c = ^dm^ which in turn implies (2.20). After Fc has been found, to determine F so that A + BF has desired eigenvalues, one should use F = FcP given in (2.13). Note that P, which reduces (A, B) to the controller form, has a specific form in this case [see (4.44) of Section 3.4]. Combining these results, it is possible to derive a formula for the eigenvalue assigning F in terms of the original pair (A, B) and the coefficients of the desired polynomial a^is). In particular, the 1 X n matrix F that assigns the n eigenvalues of A + BF at the roots of a^is) is given by F = -el%-^aM\
(2.21)
where Cn = [ 0 , . . . , 0, 1]^ G R"" and ^ = [5, A 5 , . . . , A^'-^B] is the controllability matrix. Relation (2.21) is known as Ackermann's formula and its validity will be verified shortly. Notice that the F that assigns the n eigenvalues of A + BF is unique when m = 1. This is not difficult to see from the preceding. In particular, notice that Fc = [foy' "y fn-\\ is uniquely determined by (2.19) and that the transformation matrix P is also unique in this case (m = 1) and is in fact given by (4.44) in Section 3.4. Verification of Ackermann's Formula Given (A^, Be) in controller form, it was shown that the matrix Fc = [o^Q - do,. ..,an-\
- dn-\\ = 5^^[Aj^ - Ajn]
assigns the eigenvalues of Ac + BcFc to the desired locations, the roots of a^is) given in (2.18). The matrix Fc is related to Fby F = FcP, where Ac = PAP-\ Be = PB, and P = ^c^~^ and where ^ and %c are the controllability matrices of (A, B) and (Ac, Be), respectively. These relations will now be combined to derive Ackermann's formula. First, note that if (2.21) were used for a given pair (Ac, Be) in controller form, then Fc = -el%-^aMc\
(2.22)
We now show that this Fc assigns the n eigenvalues at the desired locations. In doing so, we note that aj(Ac) = A^ + ^„_iA"~^ + • • • + d\Ac + d(^l, which in view of the Cayley-Hamilton Theorem [namely, a{Ae) = A^ + a„_iA^"^ H V aol = 0] assumes the form n-l
otd(Ac) = ^(di
- ai)A[,
i=0
since A^ = - Z f J o ^ / A J , . Also note that ef^c = e^ or that ^J^^~^ = e\. Now ~el%-^ad{Ac) = -el{{do - ao)I + ••• + (d^-i - a^-M^^) = [ao-do,..., oin-\ - dn-i] = Fe of (2.20), that is, (2.22) is verified. Note that the last relation was derived using the fact that ejAc = el, {e\ Ac) Ac = el Ac = 4 , . . . , ^f A^^'^ = e^.
from which -e^^do - ao)I H + {dn-i - an-i)A^ ^] = (o^o -
FcP=-el^-'aMc)P
which is Ackermann's formula, (2.21).
•
EX AMPLE 2.4. To the system of Example 2.3 we apply (2.21) and obtain -el^t -[0,1]
'ad{A) 2 2 L 3
2 2I] " 3 ' 33J
"17 4 9 2
-1] / W 2 2 3J 9 2
11
(
1
il
2
•1
1" 2
+ 2 21 2
"1
+2 0
0" 1
- [ 1 13] ~ L 6 ' 3 J'
which is identical to the F found in Example 2.3.
•
Now consider the multi-input case (m > 1). We proceed in a way completely analogous to the single-input case. Assume that Ac and Be are in the controller form, (4.62), given in Section 3.4. Notice that ACF = A^ + BcFc is also in (controller) companion form with identical block structure as A(^ for any FQ. In fact, the pair {AQF^BQ) has the same controllability indices /i/,/ = 1,... ,m, as {Ac^Bc). This was shown in Section 3.4 of Chapter 3 and can also be seen directly, using (4.63), since Ac + BcFc — {Ac + BcAm) + {BcBm)Fc
— Ac + Bc{Am + BmFc),
(2.23)
where Ac and Be are defined in (4.63). We can now select an n x n matrix Aj with desired characteristic polynomial det{sI — Ad) = OJj(^) =^^ + J^_i^^
-Jo,
(2.24)
and in companion form, having the same block structure as ACF or A^, that is, Aj = Ac+BcAd^. Now if AcF = Aj, then in view of (2.23), Bc{Am+BmFc) = BcA^^. From this it follows that F,=B-'[A^„-A^], (2.25) where Bm^A^^, and A^ are the m Gjth rows of 5c,Aj, and A^, respectively, and Gj = E/=i l^iij = ^i---i^' Note that this is a generalization of (2.20) of the single-input case. We shall now show how to select an n x n matrix A j in multivariable companion form to have the desired characteristic polynomial. One choice is "0 1 ••• 0 Ad
0 -Jo
0
1
-di
-dn-l
335 CHAPTER 4 :
State Feedback and State Observers
336 Linear Systems
the characteristic polynomial of which is a^{s). In this case the m x n matrix A^^ is given by ' 0 ^d^
•••
0 -do
0
1
0
0
•••
0
•••
0
0 -dn-1
where the ith row, / = l , . . . , m — l , i s zero everywhere except at the (7/ + 1 column location, where it is one. Another choice is to select A j = [Aij],iJ = 1 , . . . ,m, with Aij = 0 for / ^ 7, i.e., [All 0
0 A22
••• •••
0 0
Ad-
0
0
noting that det {si — A^) = det {si — An)""0
1
det {sI — Amm). Then •••
0"
A,
where the last row is selected so that det {si—An) has desired roots. The disadvantage of this selection is that it may impose unnecessary restrictions on the number of real eigenvalues assigned. For example, if n = 4,m = 2 and the dimensions of A n and A22, which are equal to the controllability indices, are di=3 and d2 = l, then two of the eigenvalues must be real (why?). There are of course other selections for A j , and the reader is encouraged to come up with additional choices. A point that should be quite clear by now is that Fc (or F) is not unique in the present case, since different Fc can be derived for different A j ^ , all assigning the eigenvalues at the same desired locations. In the singleinput case, Fc is unique, as was shown. Therefore, the following result has been shown. LEMMA 2.4. Let (A,5) be controllable and suppose that n desired real and complex conjugate eigenvalues for A + BF have been selected. The state feedback matrix F that assigns all eigenvalues of A + BF to desired locations is not unique in the multi-input case (m > 1). It is unique in the single-input case m= I. • EXAMPLE 2.5. Consider the controllable pair (A, 5), where 0 0 0
1 0 2
and
"0 1 0
r1 0
It was shown in Example 4.14 of Chapter 3, that this pair can be reduced to its controller form 0 2 1
PAP-
where
1 -1 0
0' 0 0.
'0 0" Be = PB = 1 1 .0 1.
"0
0
11 2
P = 0
1
1 2
.1
0
2J
1
Suppose we desire to assign the eigenvalues of A + BF to the locations {-2, - 1 ± j}, i.e., at the roots of the polynomial ad(s) = (s + 2)(s'^ + 2s + 2) = s^ + As^ + 65' + 4. A choice for Ad is 1 0 0 0 -4 -
Adi =
and
F,ci
0 1 -4.
1 0
A J
B-^Ad 1 .0
leading to A^^. =
--11 r-2
ij L-5
1 -6
1 1
r
0 .-4 0 -6
0 -6 1 -4
7 -6
-4
ively, Ad2
and
=
1 - -2 0
" 0 -2 (3
0 '-2 0 , from which Ad ^ = 0 -2
5;^[A.,-AJ -1 0 5
'1 0
- 1 1 r-4 Ij [-1
-1 0
-2 0
0' -2_
0" -2
2 -2 1 -
2 assign the and F2 = FciP = -2 -4 - 6 eigenvalues of A + 5 F to the locations { 2 , - l ± j } . The reader should plot the states of the equation x = {A + BF)x for F = Fi and F = F2 when x(0) = [1, 1, 1]^, and should comment on the differences between the trajectories. • Both Fi = FciP =
Relation (2.25) gives all feedback matrices, Fc (or F = FcP), that assign the n eigenvalues of A^ + BcFc (or A + BF) to desired locations. The freedom in selecting such Fc is expressed in terms of the different A j , all in companion form, with A^ = [Aij] and A/y of dimensions fit X fij, which have the same characteristic polynomial. Deciding which one of all the possible matrices A^ to select, so that in addition to eigenvalue assignment other objectives can be achieved, is not apparent. This flexibility in selecting F can also be expressed in terms of other parameters, where both eigenvalue and eigenvector assignment are discussed, as will now be shown. 3. Assigning eigenvalues and eigenvectors. Suppose now that F was selected so that A + BF has a desired eigenvalue Sj with corresponding eigenvector v^. Then
337 CHAPTER 4 :
State Feedback and State Observers
338
[sjl — (A + BF)]Vj = 0, which can be written as
Linear Systems [Sjl-A,B]
-Fvj_
0.
(2.26)
To determine an F that assigns Sj as a closed-loop eigenvalue, one could first determine a basis for the right kernel (null space) of [sjI — A^B], i.e., one could determine Mj a basis such that -Di Mj [Sjl-A,B] 0. (2.27) Note that the dimension of this basis is (n + m)-rank [sjl —A,B] = {n + m)—n = m, where rank [sjl —A,B] = n since the pair {A^B) is controllable. Since it is a basis, there exists a nonzero m x 1 vector Uj so that Mj-
Combining the relations —DjUj
aj =
_-Fvj_
(2.28)
—FVj and Mjaj -- Vj, one obtains FMjaj
= DjUj.
(2.29)
This is the relation that F must satisfy for Sj to be a closed-loop eigenvalue. The nonzero m x 1 vector aj can be chosen arbitrarily. Note that Mjaj = Vj is the eigenvector corresponding to Sj. Note also that aj represents the flexibility one has in selecting the corresponding eigenvector, in addition to assigning an eigenvalue. The nxl eigenvector Vj cannot be arbitrarily assigned; rather, the m x 1 vector aj can be (almost) arbitrarily selected. These mild conditions on aj are stated next as a theorem. THEOREM 2.5. The pair (sj, Vj) is an (eigenvalue, eigenvector)-pair of A + BF if and only if F satisfies (2.29) for some nonzero vector Uj such that Vj = MjUj with
Mj
a basis of the null space of [sjl —A,B\ as in (2.27). Proof, Necessity has been shown. To prove sufficiency, postmultiply Sjl — ( A + 5 F ) by Mjaj and use (2.29) to obtain (sjI — A)Mjaj —BDjaj = 0 in view of (2.27). Thus, [sjI-{A
+ BF)]Mjaj = 0.
which implies that Sj is an eigenvalue of A + BF and MjUj = Vj is the corresponding eigenvector. • If relation (2.29) is written for n desired eigenvalues Sj, where the aj are selected so that the corresponding eigenvectors Vj = Mjaj are linearly independent, then FV = W,
(2.30)
where V = [ M i ^ i , . . . ,M^a^] and W = [ D i ^ i , . . .^Dnan] uniquely specify F as the solution to these n linearly independent equations. When Sj are distinct, it is always possible to select aj so that V has full rank; in fact almost any set of nonzero aj suffices. When Sj have repeated values it may still be possible under certain conditions to select aj so that Mjaj are linearly independent; however in general, for multiple eigenvalues, (2.30) needs to be modified, and the details for this can be found in the literature (e.g., [16]).
It can be shown that for distinct eigenvalues Sj, the n vectors Mjaj, j = 1,.. .,n, are Hnearly independent for almost any nonzero aj. For further discussion on this, see the remarks following Example 2.6 and Subsection A.4, on Interpolation, in the Appendix. Also note that if Sj+i = s*, the complex conjugate of Sj, then the corresponding eigenvector V/+1 = v* = ^)^)' Relation (2.30) clearly shows that the F that assigns all n closed-loop eigenvalues is not unique (see also Lemma 2.4). All such F are parameterized by the vectors aj that in turn characterize the corresponding eigenvectors. If the corresponding eigenvectors have been decided upon—of course within the set of possible eigenvectors Vj = MjUj—then F is uniquely specified. Note that in the single-input case, (2.29) becomes FMj = Dj, where Vj = Mj. In this case, F is unique. EXAMPLE 2.6. Consider the controllable pair (A, B) of Example 2.5 given by
11 1 1 .0 0.
0 1 1 0 2 -1
0 0 0
ro
Again, it is desired to assign the eigenvalues of A + BF at - 2 , - 1 ± j . Let ^-i = -2,S2 = -I + j and S3 = -I - j . Then, in view of (2.27),
Ml]
1 -1 2
=
'
M2 -D2.
M3 -D3
Ml
1 0 0
2+ J 1
-1 + 7 1-J
=
-1 -2 2_ 1 and
J 2
r
0 0
, the complex conjugate, since S3
Each eigenvector v^ = Miai, i = 1, 2, 3, is a linear combination of the columns of Mi. Note that V3 = V2. If we select the eigenvectors to be '1 y = [vi, V2, V3] = 0 0 I.e., ai =
a2
, and a3 1 0 0
1 j 2
1 j 2
r
then (2.30) implies that -2-j -1
-
-2 + j -1
from which we have -2-j -1
'^rj 2 -2
-1 0
-2 + J -1
4y 0 0
0 2 -2
-2/ j j .
-2 ^
This matrix F is such that A + BF has the desired eigenvalues and eigenvectors (verify this). •
339 CHAPTER 4: State Feedback and State Observers
340
Remarks
Linear Systems
At this point, several comments are in order. 1. In Example 2.6, if the eigenvectors were chosen to be the eigenvectors of A + BFi (instead of A + BF2) of Example 2.5, then from FV = W it follows that F would have been Fi (instead of F2). 2. When st = ^*^j, then the corresponding eigenvectors are also complex conjugates, i.e., Vi = v*^j. In this case we obtain from (2.30) that FV = Fl..,
ViR + pii, ViR -
jvii,...]
= [..., WiR + jwii, WiR - jwii, ...] = W.
Although these calculations could be performed over the complex numbers (as was done in the example), this is not necessary, since postmultiplication of FV = Why
-J', +7
shows that the above equation FV = W is equivalent to Fl..,
ViR, Vii,...]
= [..., WiR,
Wii,...],
which involves only reals. 3. The bases
U j = 1, • • •, n, in (2.27) can be determined in an alternative way -Dj\' and the calculations ons can be simplified if the tl controller form of the pair (A, B) is
known. In particular, note that [si - A,B]\
D{s)
= 0, where the n X m ma-
trix S(s) is given by S(s) = block diag [1, >y,..., 5^'"^] and the ^c^, / = 1 , . . . , m, are the controllability indices of (A, B). Also, the m X m matrix D{s) is given by D{s) = B^^ [diag [s^^\ ..., s^"'] - AmS{s)l Note that S{s) and D{s) were defined in the Structure Theorem (controllable version) in Chapter 3 (Subsection 3.4D). It was shown there that {si - Ac)S{s) = BcD(s), from which it follows that (si A)P~^S(s) = BD(s), where P is a similarity transformation matrix that reduces (A, B) to the controller form (Ac = PAP-\ Be = PB). Since P~^S(s) and D(s) are right coprime polynomial matrices (see Section 7.2 of Chapter 7), we have rank
D(sj)
= m for any Sj, and therefore,
P-'S{Sj) D(sj)
qualifies as a basis
for the null space of the matrix [sjl - A, B](P = I when A, B are in controller
form, i.e., A = Ac and B = Be.) We note that this approach is discussed further in Subsection A.4 of the Appendix. Returning to Example 2.6, the controller form of (A, B) was found in Example 2.5 using
p-'
Here
"1 0 r — 1 1 0 2 0 0
'\ 0" s 0 0 1
S(s)
^2
D{s) = '-
-s s 1 0 0
1 s+1 2
M{s) -Dis)
and
+ s- 1 -1
-(-$2 + 5 - 1 )
s -s
Then
Ml]
'
-Di\
Mi-2) -Di-2)_
1 -1 2
r
0 0
-1 -2 1 2_
Mi-l -D(-l
M2 -£>2
+ j) + j)
1 0 0
J 2 2+ j 1
and
Ms
1+ ; 1-7
which are precisely the bases used in the example.
4. If in Example 2.6 the only requirement were that (^'i, vi) = ( - 2 , (1, 0, 0)^), then F(h 0, Of = (2, - 2 f , i.e., any F = \l {'^ ^}' will assign the desired values L^/22 723. to an eigenvalue of A + BF and its corresponding eigenvector. 5. All possible eigenvectors vi and V2(v3 = v^) in Example 2.6 are given by
1] |- -, 0 an and V2 = M2a2 2 Oj [ai2_ 1
v\ = M\a\ = - 1
/I
ail + JC131 [a22 + JC132
where the atj are such that the set {vi, V2, V3} is linearly independent (i.e., y = [vi, V2, V3] is nonsingular) but otherwise arbitrary. Note that in this case (sj
341 CHAPTER 4 :
State Feedback and State Observers
342 Linear Systems
distinct), almost any arbitrary choice for atj will satisfy the above requirement (see Appendix, Subsection A.4). C. The Linear Quadratic Regulator (LQR): Continuous-Time Case A linear state feedback control law that is optimal in some sense can be determined as a solution to the so-called Linear Quadratic Regulator (LQR) problem (more recently, also called the H2 optimal control problem). The LQR problem has been studied extensively, and the interested reader should consult the extensive literature on optimal control for additional information on the subject. In the following, we give a brief outline of certain central results of this topic to emphasize the fact that the state feedback gain F can be determined to satisfy, in an optimal fashion, requirements other than eigenvalue assignment, discussed above. The LQR problem has been studied for the time-varying and time-invariant cases. Presently, we will concentrate on the time-invariant optimal regulator problem. Consider the time-invariant linear system given by X = Ax-\- Bu,
z = Mx,
(2.31)
where the vector z(t) represents the variables to be regulated—to be driven to zero. We wish to determine u{t), t ^ 0, which minimizes the quadratic cost J(u) =
[z' (t)Qz(t) + u' (t)Ru(t)] dt (2.32) Jo for any initial state x{0). The weighting matrices Q, R are real, symmetric, and positive definite, i.e., Q = Q^, R = R^, and Q> 0, R > 0. This is the most common version of the LQR problem. The term z^Qz = x^ {M^ QM)x is nonnegative, and minimizing its integral forces z{t) to approach zero as t goes to infinity. The matrix M^QM is in general positive semidefinite, which allows some of the states to be treated as "do not care" states. The term u^Ru with R > 0 is always positive for w 7^ 0, and minimizing its integral forces u(t) to remain small. The relative "size" of Q and R enforces tradeoffs between the size of the control action and the speed of response. Assume that (A, B, Q^'^^M) is controllable (-from-the-origin) and observable. It turns out that the solution w*(0 to this optimal control problem can be expressed in state feedback form, which is independent of the initial condition x(0). In particular, the optimal control w* is given by u{t) - F*x(0 = -R~^B^Plx{t\
(2.33)
where P* denotes the symmetric positive-definite solution of the algebraic Riccati equation A^Pc + PcA - PcBR'^B^Pc
+ M^QM = 0.
(2.34)
This equation may have more than one solution, but only one that is positive-definite (see Example 2.7). It can be shown that w*(0 "= F*x(t) is a stabilizing feedback control law and that the minimum cost is given by Jniin = /(w*) = x^(0)P*x(0), The assumptions that (A, B, Q^'^^M) are controllable and observable may be relaxed somewhat. If (A, B, Q^'-^M) is stabilizable and detectable, then the uncon-
trollable and unobservable eigenvalues, respectively, are stable, and P* is the unique, symmetric, but now positive-semidefinite solution of the algebraic Riccati equation. The matrix F* is still a stabilizing gain but it is understood that the uncontrollable and unobservable (but stable) eigenvalues will not be affected by F*. Note that if the time interval of interest in the evaluation of the cost goes from 0 to ^1 < 00, instead of 0 to oo, that is, if J(u) =
[z^(t)Qz(t) + u^(t)Ru(t)] dt,
(2.35)
then the optimal control law is time-varying and is given by u{t) = -R~^B^P\t)x{t),
Q^t^ti,
(2.36)
where P*(0 is the unique, symmetric, and positive-semidefinite solution of the Riccati equation, which is a matrix differential equation of the form - — P{t) = A^P(t) + P(t)A - P(t)BR-^B^P(t) dt
+ M^QM,
(2.37)
where P(ti) = 0. It is interesting to note that if (A, B, Q^'^M) is stabilizable and detectable (or controllable and observable), then the solution to this problem as ti —> 00 approaches the steady-state value P* given by the algebraic Riccati equation; that is, when ti^ ^ the optimal control policy is the time-invariant control law (2.33), which is much easier to implement than time-varying control policies. EXAMPLE 2.7. Consider the system described by the equations i = Ax+Bu,y = Cx, "0 r ,B = "0" C = [1, 0]. Then (A, B, C) is controllable and observable where A 0 0. .1_ and C{sl -A) ^B = \ls^. We wish to determine the optimal control M*(0, ^ ^ 0, which minimizes the performance index (/(O + pu^iO) dt,
J=
where p is positive and real. Then R = p> 0, zit) = y(t), M = C, and Q the present case the algebraic Riccati equation (2.34) assumes the form
1 > 0. In
A^Pc + PcA - PcBR-^B^Pc + M^QM 0 0 1 0 0 0 1 0
Pi P2
Pi P3
0 0
where Pr =
Pi P2
1
Pr + Pr
0
p
Pi IP2
P2 P3.
[0 l]Pc + 1 Pi PIP2
Oj
[10]
P2]ro 0] \pi P3\
[0
P2 [P2 P3 ij
0 0
= P^. This implies that
IP2 P3j
1 1 ~-P2 + 1 = 0 , pi - -P2P3 = 0, 2p2 :PI = 0. P P P Now Pc is positive definite if and only if pi > 0 and p\p3 — p\> 0. The first equation above implies that p2 = ± J^. However, the third equation, which yields p\ = 2pp2,
343 CHAPTER 4 :
State Feedback and State Observers
344 Linear Systems
implies that pi = + yp. Then p\ = 2p J~p and p^ = ±
2p J^. The second equation
yields pi = {\lp)p2P?> and implies that only p^ = + lip J~p is acceptable, since we must have p\ > 0 for P^ to be positive definite. Note that pi > 0 and P3-p\ = 2p-p = p > 0, which shows that >
VP
J^P/P.
is the positive definite solution of the algebraic Riccati equation. The optimal control law is now given by u\t) = F*x(t) = -R'^B^PlxQ)
= - - [ 0 , IjPXO.
The eigenvalues of the compensated system, i.e., the eigenvalues of A + BF*, can now be determined for different p. Also, the corresponding u*(t) and y{t) for given x(0) can be plotted. As p increases, the control energy expended to drive the output to zero is forced to decrease. The reader is asked to verify this by plotting u*(t) and y(t) for different values of p when x(0) = [1, 1]^. Also, the reader is asked to plot the eigenvalues of A + BF* as a function of p and to comment on the results. • It should be pointed out that the locations of the closed-loop eigenvalues, as the weights Q and R vary, have been studied extensively. Briefly, for the singleinput case and for Q = ql and R = r in (2.32), it can be shown that the optimal closed-loop eigenvalues are the stable zeros of 1 + (q/r)H^(-s)H(s), where H(s) = M(sl - Ay^B. As q/r varies from zero (no state weighting) to infinity (no control weighting), the optimal closed-loop eigenvalues move from the stable poles of H'^(-s)H(s) to the stable zeros of H'^(-s)H(s). Note that the stable poles of H^(-s)H(s) are the stable poles of H(s) and the stable reflections of its unstable poles with respect to the imaginary axis in the complex plane, while its stable zeros are the stable zeros of H(s) and the stable reflections of its unstable zeros. The solution of the LQR problem relies on solving the Riccati equation. A number of numerically stable algorithms exist for solving the algebraic Riccati equation. The reader is encouraged to consult the literature for computer software packages that implement these methods. A rather straightforward method for determining P* is to use the Hamiltonian matrix given by H^
A -M^QM
-BR-^B^ -A^
(2.38)
Let [V\, V2V denote the n eigenvectors of H that correspond to the n stable [Re (A) < 0] eigenvalues. Note that of the 2n eigenvalues of H, n are stable and are the mirror images reflected on the imaginary axis of its n unstable eigenvalues. When (A, B, Q^''^M) is controllable and observable, then H has no eigenvalues on the imaginary axis [Re (A) = 0]. In this case the n stable eigenvalues of H are in fact the closed-loop eigenvalues of the optimally controlled system, and the solution to the algebraic Riccati equation is then given by P* = V^V^K
(2.39)
Note that in this case the matrix Vi consists of the n eigenvectors of A + i5F*, since for Ai a stable eigenvalue of H, and vi the corresponding (first) column of Vi, we have [Ai/ - (A + 5F*)]vi = [Ai/ - A + BR-^B^ V2yr^]vi [Ai/,0] 0
X
0
X
0
X
0
X
Vi'v,
{A,-BR-'B''[
Vx vi
where the fact that
Vi
are eigenvectors of H was used. It is worth reflecting for a
moment on the relationship between (2.39) and (2.30). The optimal control F derived by (2.39) is in the class of F derived by (2.30). D. Input-Output Relations It is useful to derive the input-output relations for a closed-loop system that is compensated by linear state feedback, and several are derived in this subsection. Given the uncompensated or open-loop system x = Ax + Bu, y = Cx + Du, with initial conditions x(0) = XQ, we have y(s) = C(sl - Ay^xo + H(s)u(sl
(2.40)
where the open-loop transfer function H(s) = C(sI—A)~^B-\-D. Under the feedback control law u = Fx -\- r, the compensated closed-loop system is described by the equations x = (A + BF)x -h 5r, j = (C + DF)x + Dr, from which we obtain y{s) = (C + DF)[sI - (A + BF)r^xo
+ HF(s)r{sl
(2.41)
where the closed-loop transfer function Hf(s) is given by HF(S) = ( C + DF)[sI - (A + BF)Y^B -f- D. Alternative expressions for HF{S) can be derived rather easily by substituting (2.4), namely, u{s) = F[sl - (A + BF)r^xo
+ [/ - F(sl -
Ay^Br^r(sl
into (2.40). This corresponds to working with an open-loop control law that nominally produces the same results when applied to the system [see the discussion on open- and closed-loop control that follows (2.4)]. Substituting, we obtain y{s) = [C(sl - A)"i + H(s)F[sI - (A -t- BF)r^]xo + H(s)[I - F(sl -
Ay^Br^r(s).
(2.42)
345 CHAPTER 4 :
State Feedback and State Observers
346 Linear Systems
Comparing with (2.41), we see that (C + DF){sI - (A + BF)Y^ H(s)F[sI - (A + BF)]~\ and that HF(S) = (C + DF)[sI - (A + BF)r^B
= C(sl - A)~^ +
+ D
= [C(sl - Ay^B + D][I - F(sl -
Ay^B]-^
= H(s)[I - F(sl - Ay^Br^^
(2.43)
The last relation points out the fact that y(s) = HF(s)r(s) can be obtained from y(s) = H(s)u(s) using the open-loop control u(s) = [/ - F{sl - Ay^B]~^r(s). Relation (2.43) can easily be derived in an alternative manner, using fractional matrix descriptions for the transfer function, introduced in Section 3.4 (see the Structure Theorem). In particular, the transfer function H{s) of the open-loop system {A, B, C, D} is given by H(s) -
N(s)D-\s\
where A^(^) = CS(s) + DD{s) with S(s) and D(s) satisfying (si - A)S(s) = BD(s) (refer to the proof of the controllable version of the Structure Theorem given in Section 3.4). Notice that it has been assumed, without loss of generality, that the pair (A, B) is in controller form. Similarly, the transfer function HF(S) of the compensated system {A+BF, B,C+ DF, D] is given by HF{S)
=
NF(S)D^\S\
where NF(S) = (C + DF)S(s) + DDF(S) with S(s) and DF(S) satisfying [si - (A + BF)]S(s) = BDF(S). This relation implies that (si - A)S(s) = B[DF(S) + FS(s)l from which we obtain DF(S) + FS(s) = D(s). Then NF(S) = CS(s) + D[FS(s) + DF(S)] = CS(s) + DD(s) = N(s\ that is, HF(S) = N(s)Dp\s\
(2.44)
where DF(S) = D(s) - FS(s). The full justification of the validity of these expressions will be given in Subsection 7.4B of Chapter 7, where feedback systems described in terms of polynomial matrix representations are addressed. Note that / - F(sl - Ay ^B in (2A3) is the transfer function of the system {A,B, - F , / } and can be expressed as D/7(5')D"^(^), where D/7(^) = -FS(s)+ID(s). LetM(^) = (Z)/7(^)^~k^))"^. Then (2.44) assumes the form HF(S) = N(s)Dp\s)
= (N(s)D-\s))(D(s)D^\s))
= H(s)M(s).
(2.45)
Note that relation HF(S) = N(s)D^^(s) also shows that the zeros of H(s) [in N(s), see also Subsection 7.3B] are invariant under linear state feedback; they can be changed only via cancellations with poles. Also observe that M(s) = D(s)Df^(s) is the transfer function of the system {A + BF, B, F, I] (show this). This implies that HF(S) in (2.43) can also be written as HF(S) = H(s)[F(sI - (A + BF)y^B
+ /],
a result that could also be shown directly using matrix identities. EXAMPLE 2.8. Consider the system x = Ax + Bu, y = Cx, where 0 2 1
1 -1 0
0 0 0
and
B =^ Be
ro 0] 1 .0
1 1.
as in Example 2.5, and let C = Q = [1, 1, 0]. Hf(s) will now be determined. In view of the Structure Theorem developed in Section 3.4, the transfer function is given by H(s) = N(s)D-\sX where
n N(s) = CcS(s) = [1,1,0] s 0 and
D{s)=^ B-'[A{s)-AmS(s)] - 1 s^ -\-s-2 -1 1
1 0
1 0
= 0 s
01 0 = [^ + 1,0], 1. s^ 0
1 1
ri 01" "2 - 1 01 \s 0 1 0 Oj Lo i j .
0 s
s^ + s- I
Then H{s) = N(s)D-'is)
= [s+ 1,0]
[s + 1, 0] s 1
s'^ + s- 1 -1
s s^ +
S-1 ^3 + 52 - 25
1 [5(5+1), 5(5+1)] = - ^ ± i - [ l , l ] . s(s^ + 5 - 2 ) ' ' '' ' '' 52 + 5 - 2 3 7 5 (which is Fci of Example 2.5), then -5 -6 -4
IfF, =
5^ + 5 — 1 —5 -1
D^(5) = Z)(5) - FcS(s) = 52 - 65 - 4 65 + 4
3 7 -5 -6
ri 0" 51 5 0 -4j
Lo 1 .
-5-5 5+ 4
Note that detDpis) = 5^ + 452 + 65 + 4 = (5 + 2)(52 + 25 + 2) with roots - 2 , - 1 ± 7, as expected. Now HF{S) = N{s)D}\s)
= [5 + 1,0]
5+4 -65 - 4
1 5+5 5 2 - 6 5 - 4 (5 + 2)(52 + 25 + 2)
5+ 1 [5 + 4, 5 + 5]. (5 + 2)(52 + 25 + 2) Note that the zeros of H{s) and Hp{s) are identical, located at - 1 . Then Hpis) H(s)M(s), where M(s) = D(s)Dp\s)
=
52 + 5 - 1
5^ + I I 5 2 + 7 5 - 4 -652 - 55 - 4
=
5+4 -65-4
-
1 5+5 5 2 - 6 5 - 4 53 +452 + 65 + 4
1 1252 + 8 5 - 5 6 5 2 - 5 5 - 5 53 +452 + 65 + 4
[I~F,(sI-Acr'Bc]-K
Note that the open-loop uncompensated system is unobservable, with 0 being the unobservable eigenvalue (why?), while the closed-loop system is observable, i.e., the control law changed the observability of the system. •
347 CHAPTER 4 :
State Feedback and State Observers
348 Linear Systems
E. Discrete-Time Systems Linear state feedback control for discrete-time systems is defined in a way that is analogous to the continuous-time case. The definitions are included here for purposes of completeness. We consider a linear, time-invariant, discrete-time system described by equations of the form x(k + 1) - Ax(k) + Bu(kl y(k) = Cx(k) + Du(k),
(2.46)
where A G /?"><^ B E Z^^^^, C G /?^><^ D G 7?^x^, and k > ko, with k ^ ko = 0 being typical (see Section 2.7). DEFINITION2.3. The linear (discrete-time, time-invariant) state feedback control law is defined by u(k) = Fx(k) + r(k), where F G T?'"^" is a gain matrix and r(k) G R^ is the external input vector.
(2.47) •
This definition is similar to Definition 2.1 for the continuous-time case. The compensated closed-loop system is now given by x(k + 1) - (A 4- BF)x(k) + Br(k) y(k) = (C + DF)x(k) + Dr(kl
(2.48)
In view of Section 2.7 of Chapter 2, the system x(k + 1) = (A + BF)x(k) is asymptotically stable if and only if the eigenvalues of A + BF satisfy |A/| < 1, i.e., if they lie strictly within the unit disc of the complex plane. The stabilization problem for the time-invariant case therefore becomes a problem of shifting the eigenvalues of A + BF, which is precisely the problem studied before for the continuous-time case. Theorem 2.1 and Lemmas 2.2 and 2.3 apply without change, and the methods developed before for eigenvalue assignment can be used here as well. The only difference in this case is the location of the desired eigenvalues: they are assigned to be within the unit circle to achieve stability. We will not repeat here the details for these results. Input-output relations for discrete-time systems, which are in the spirit of the results developed in the preceding subsection for continuous-time systems, can be derived in a similar fashion, this time making use of the z-transform of x(k + 1) = Ax(k) + Bu(k), x(0) = XQ to obtain x(z) = z(zl - Ay^xo + (zl - Ar^Bu(z).
(2.49)
[Compare expression (2.49) with x(^) = (si- A)~^xo-^(sI-A)~^Bu(s).] The reader is asked to derive formulas for the discrete-time case that are analogs to expressions (2.40) to (2.45) for the continuous-time case.
F. The Linear Quadratic Regulator (LQR): Discrete-Time Case The formulation of the LQR problem in the discrete-time case is analogous to the continuous-time LQR problem. Consider the time-invariant linear system x(k + i) = Ax(k) + Bu(kl z(k) = Mx(k\
(2.50)
where the vector z(t) represents the variables to be regulated. The LQR problem is to determine a control sequence {u*(k)}, /: > 0, which minimizes the cost function J(u) = ^[z^(k)Qz(k)
+ u^(k)Ru(k)]
(2.51)
k=0
for any initial state x(0), where the weighting matrices Q and R are real symmetric and positive definite. Assume that (A, B, Q^'^M) is reachable and observable. Then the solution to the LQR problem is given by the linear state feedback control law u{k) = F*x(^) - -[R + B^PlBr^B^PlAx(k),
(2.52)
where P* is the unique, symmetric, and positive-definite solution of the (discretetime) algebraic Riccati equation, given by Pc = A'^lPc - PcB[R + B^PcBr^B^PM
+ M^QM.
(2.53)
Theminimum valueof/is/(w*) = /min = x^(0)P*x(0). As in the continuous-time case, it can be shown that the solution P* can be determined from the eigenvectors of the Hamiltonian matrix, which in this case is H
A+
BR-^B^A-^M^QM -A-^M^QM
-BR-^B^A-^ A-^
(2.54)
where it is assumed that A~^ exists. Variations of the above method that relax this assumption exist and can be found in the literature. Let [V\, Vj]^ be n eigenvectors corresponding to the n stable (|A| < 1) eigenvalues of H. Note that out of the In eigenvalues of //, n of them are stable (i.e., within the unit circle) and are the reciprocals of the remaining n unstable eigenvalues (located outside the unit circle). When (A, B, Q^'^M) is controllable (-from-the-origin) and observable, then if has no eigenvalues on the unit circle (|A| = 1). In fact the n stable eigenvalues of// are in this case the closed-loop eigenvalues of the optimally controlled system. The solution to the algebraic Riccati equation is given by -1
K - ^2vr
(2.55)
As in the continuous-time case, we note that Vi consists of the n eigenvectors of A + 5F* (show this). EXAMPLE 2.9. We consider the system x{k + 1) = Ax{k) + Bu{k), y(k) = Cx(k), "0" "0 1" ,B = C = [1, 0] and we wish to determine the optimal control where A = .0 0^ .1_ sequence {u*(k)}, /: > 0, that minimizes the performance index J(u) = Y.(y\k)
+ puHk)),
k=0
where p > 0. In (2.51), z(k) = y(k), M = C,Q = 1, and /? = p. The reader is asked to determine u'^k) given in (2.52) by solving the discrete-time algebraic Riccati equation (2.53) in a manner analogous to the solution in Example 2.7 (for the continuous-time algebraic Riccati equation). •
349 CHAPTER 4 :
State Feedback and State Observers
350 Linear Systems
4.3 LINEAR STATE OBSERVERS Since the states of a system contain a great deal of useful information, there are many applications where knowledge of the state vector over some time interval is desirable. It may be possible to measure states of a system by appropriately positioned sensors. This was in fact assumed in the previous section, where the state values were multiplied by appropriate gains and then fed back to the system in the state feedback control law. Frequently, however, it may be either impossible or simply impractical to obtain measurements for all states. In particular, some of the states may not be available for measurement at all (as in the case, for example, with temperatures and pressures in inaccessible parts of a jet engine). There are also cases where it may be impractical to obtain state measurements from otherwise available states because of economic reasons (e.g., some of the sensors may be too expensive) or because of technical reasons (e.g., the environment may be too noisy for any useful measurements). Thus, there is a need to be able to estimate the values of the state of a system from available measurements, typically outputs and inputs (see Fig. 4.3). Given the system parameters A, B, C, D and the values of the inputs and outputs over a time interval, it is possible to estimate the state when the system is observable. This problem, a problem in state estimation, is discussed in this section. In particular, we will address at length the so-called full-order and reduced-order asymptotic estimators, which are also called full-order and reduced-order observers.
u
y [A, B, C, D}
f
1
state observer
•^
FIGURE 4.3
A. Full-Order Observers: Continuous-Time Systems We consider systems described by equations of the form X = Ax + Bu, where A G /^"><^ B G T^^^^, C G
RP''''
y = Cx + Du, and D G
(3.1)
RP'''^.
Full-state observers: The identity observer An estimator of the full state x(t) can be constructed in the following manner. We consider the system 'x = Ax + Bu^K(y-
yl
(3.2)
351
^ A
where y = Cx -\- Du. Note that (3.2) can be written as
CHAPTER 4 :
'x = {A- KC)x +{B-
KD, K]
(3.3)
which clearly reveals the role of u and y (see Fig. 4.4). The error between the actual state x{t) and the estimated state x{t), e(t) = x(t) - x(t), is governed by the differential equation e(t) = x(t) - jc(t) = [Ax + Bu] - [Ax + Bu + KC(x - x)] or
e(t) = [A-
(3.4)
KC]e(tl
Solving (3.4), we obtain e(t) = exp [(A
KC)t]e(0).
(3.5)
Now if the eigenvalues of A - KC are in the left half-plane, then e(t) -> 0 as ^ ^ 00, independently of the initial condition ^(0) = x(0) - x(0). This asymptotic state estimator is known as the Luenberger observer.
Li
B-KD
:§>Ht) A-KC
FIGURE 4.4 LEMMA 3.1. There exists K E R'^^P SO that the eigenvalues of A - KC are assigned to arbitrary real or complex conjugate locations if and only if (A, C) is observable. Proof, The eigenvalues of (A - KCY = A^ - C^K^ are arbitrarily assigned via K^ if and only if the pair (A^, C^) is controllable (see Theorem 2.1 of the previous section), or equivalently, if and only if the pair (A, C) is observable. • If (A, C) is not observable, but the unobservable eigenvalues are stable, i.e., (A, C) is detectable, then the error e{t) will still tend to zero asymptotically. However, the unobservable eigenvalues will appear in this case as eigenvalues of A - KC (show this), and they may affect the speed of the response of the estimator in an undesirable way. For example, if the unobservable eigenvalues are stable but are located close to the imaginary axis, then their corresponding modes will tend to dominate the response, most likely resulting in a state estimator that converges too slowly to the actual value of the state. Where should the eigenvalues of A - KC be located? This problem is dual to the problem of closed-loop eigenvalue placement via state feedback and is equally difficult to resolve. On one hand, the observer must estimate the state sufficiently fast, which implies that the eigenvalues should be placed sufficiently far from the imaginary axis so that the error e{t) will tend to zero sufficiently fast. On the other hand, this requirement may result in a high gain K, which tends to amplify exist-
State Feedback and State Observers
352 Linear Systems
ing noise, thus reducing the accuracy of the estimate. Note that in this case, noise i^ the only Hmiting factor of how fast an estimator may be, since the gain K is realized by an algorithm and is typically implemented by means of a digital computer. Therefore, gains of any size can easily be introduced. Compare this situation with the limiting factors in the control case, which is imposed by the magnitude of the required control action (and the limits of the corresponding actuator). Typically, the faster the compensated system, the larger the required control magnitude. One may of course balance the trade-offs between speed of response of the estimator and effects of noise by formulating an optimal estimation problem to derive the best K. For this, one commonly assumes certain probabilistic properties for the process. Typically, the measurement noise and the initial condition of the plant are assumed to be Gaussian random variables, and one tries to minimize a quadratic performance index. This problem is typically referred to as the Linear Quadratic Gaussian (LQG) estimation problem. This optimal estimation or filtering problem can be seen to be the dual of the quadratic optimal control problem of the previous section, a fact that will be exploited in deriving its solution. Note that the well-known Kalman filter is such an estimator. In the following, we shall briefly discuss the optimal estimation problem. First, however, we shall address the following related issues. 1. Is it possible to take ^ = Ointheestimator (3.2)? Such a choice would eliminate the information contained in the term y — y from the estimator, which would now be of the form X = A i + Bu.
(3.6)
In this case the estimator would operate without receiving any information on how accurate the estimate x actually is. The error e{t) = x(t) — x(t) would go to zero only when A is stable. There is no mechanism to affect the speed by which x(t) would approach x(t) in this case, and this is undesirable. One could perhaps determine x(0), using the methods in Section 3.3 of Chapter 3, assuming that the system is observable. Then, by setting x(0) = x(0), presumably x(t) = x(t) for all t > 0, in view of (3.6). This of course is not practical for several reasons. First, the calculated x(0) is never exactly equal to the actual x(0), which implies that ^(0) would be nonzero. Therefore, the method would rely again on A being stable, as before, with the advantage here that ^(0) would be small in some sense and so e(t) -> 0 faster. Second, this scheme assumes that sufficient data have been collected in advance to determine (an approximation to) x(0) and to initialize the estimator, which may not be possible. Third, it is assumed that this initialization process is repeated whenever the estimator is restarted, which may be impractical. 2. If derivatives of the inputs and outputs are available, then the state x(t) may be determined directly (see Exercise 3.14 in Chapter 3). The estimate x(t) is in this case produced instantaneously from the values of the inputs and outputs and their derivatives. Under these circumstances, x(t) is the output of a static state estimator, as opposed to the above dynamic state estimator, which leads to a state estimate x(t) that only approaches the actual state x(t) asymptotically as r -^ 00 [e(t) = x(t) - x(t) -^ 0 as r ^ oo]. Unfortunately, this approach is in general not viable since noise present in the measurements of u(t) and y(t) makes accurate calculations of the derivatives problematic, and since errors in u(t), y(t)
and their derivatives are not smoothed by the algebraic equations of the static estimator (as opposed to the smoothing effects introduced by integration in dynamic systems). It follows that in this case the state estimates may be erroneous.
M P L E 3.1. Consider the observablepair "0 1 0^ A= 0 0 1 ' .0 2 - 1 _
C == [1,0,0].
We wish to assign the eigenvalues of A - ^ C in a manner that enables us to design a full-order/full-state asymptotic observer. Let the desired characteristic polynomial be a J (5*) = s^ + dis^ + d\s + do and consider AD ^
^
AT
0 1 0
0 0 1
0 2 -1
and
Bn = C T _
To reduce (AD, BD) to controller form, we consider %
Then
[BD,ADBD,AIBD]
ro
0
1"
P = UAD
0
1
-1
IqAl
Ll
-1
3.
from which we obtain
AD,
=
PADP~
0 0 0
"1 0 0" = 0 1 0 = ^ .0 0 1. and
1 0 2
-2 1 1
P-^ =
and
BD,
1 1 1 0 0 0
= PBE
The state feedback is then given by FD, = B^^[Ad^ - Am] = [-do, -di - 2, -d2 + I] and FD = FDCP = [~d2 + I, d2 — d\ — 3, d\ — do — 3d2 + 5]. Then
K = -Fl
= [d2 -hdi-d2
+ 3, do-di+
M2 - 5]^
assigns the eigenvalues of A - KC at the roots of 0Ld{s) = s^ + d2S^ + d\s + do. Note that the same result could also have been derived using the direct method for eigenvalue assignment, using |^/ - (A - (^0, h, fe)^C)| = ad{s). Also, the result could have been derived using the observable version of Ackermann's formula, namely, K = -Fl=
ad(A)€~'en,
where FD = -el^^^ad(AD) from (2.21). Note that the given system has eigenvalues at 0,1, - 2 and is therefore unstable. The observer derived in this case will be used in the next section (Example 4.1) in combination with state feedback to stabilize the system X = Ax + Bu, y = Cx, where 0 0 0
1 0 2
0' 1 -1.
B =
"0 1 1 1 .0 0
and
C = [1, 0, 0]
(see Example 2.5), using only output measurements. EXAMPLE 3.2. Consider the system i: = Ax,y = Cx, where A =
andC =
[0,1], and where (A, C) is in observer form. It is easy to show that K = [do- 2,
d\-2Y
353 CHAPTER 4 :
State Feedback and State Observers
354 Linear Systems
assigns the eigenvalues of A - KC at the roots of s^ + d\s + d^. To verify this, note that det{sI-{A-KC))
= det\ s 0 0 s
0 .1
-do -d\\
s + d\s + d{).
The error e{t) = x(t) - x(t) is governed by the equation e(t) = (A - KC)e(t) given in (3.4). Noting that the eigenvalues of A are - 1 ± j , select different sets of eigenvalues for the observer and plot the states x(t), x(t) and the error e(t) for x(0) = [2, 2]^ and x(0) = [0, 0]^. The further away the eigenvalues of the observer are selected from the imaginary axis (with negative real parts), the larger the gains in K will become and the faster x(t) -^ x(t) (verify this). • Partial or linear functional state observers The state estimator studied above is a full-state estimator or observer, i.e., x(t) is an estimate of the full-state vector x(t). There are cases where only part of the state vector, or a linear combination of the states, is of interest. In control problems, for example, Fx(t) is used and fed back, instead of Fx(t), where F is an m X ^ state feedback gain matrix (see also Section 4.4). An interesting question that arises at this point is: is it possible to estimate directly a linear combination of the state, say, Tx, where T E R^'^^^ h^ nl This problem is considered next. We consider the system z = Az + Bu + Ky,
(3.7)
where A G R^^"", B G /^^x^, and K G R^'^P are to be determined. The error equation is given by ^ = Tx-
z = TAx + TBu - Az-
Bu-
K{Cx + Du)
= Ae + {TA - AT - KC)x + {TB - KD - B)u. Now if B = TB-
KD
(3.8)
and r . A, and K satisfy TA-
AT = KC,
(3.9)
then e = Ae. If in addition, A is stable, then e{t) -^ 0 as t ^ oo and z(t) will approach Tx(t) asymptotically. A key issue is clearly the existence (and calculation) of appropriate solutions of (3.9). This has been studied extensively in the literature (see for example O'Reilly [18]). Here we simply wish to point to a special case of (3.9), namely, the case of the identity observer (T = I) or of the full-state estimator: T = I and A = A - KC, where for (A, C) observable, there always exists K that renders A stable, as was shown above. Note that in general a solution (A, K) of (3.9) with A stable may not exist, i.e., there may not exist an observer (3.7) of order n(A G R^^^) to estimate n linear functions of the state, Tx (T G R^^^). It is possible, however, to decouple the order of the observer from the number of linear functions of the state to be estimated, in the following manner. Let w = Tx, w ^ R^, and consider z = Az-^ Bu-\- Ky w = Tx{y-Du)
+ T2Z,
(3.10)
where z G i?^ A G R'""', B G i^'^^, K G i?'^^^^, Ti G /?^><^, and 72 G R^""'. Let f G W"" and write
+ KD)u.
Now if B = TB - KD mdKC - TA = -AT, where A is stable, then z-fx = A(z - tx) or z - t x = e^\z(0) - fx(0)), i.e., z(t) -^ fx(t). We are interested, however, in estimates of w = Tx. Consider therefore w — w = [TiCx + Tiifx + e^\z{Qi) - fx(0))] -Tx = {TiC + T2f - T)x + T2e^\z{0) - fx(0)). Now if T TiC + T2f, then w-w = T2e^\z{^) - fx(0)) and w(t) -^ w(t) = Tx{t\ since A is stable. Therefore, an observer (3.10) of w = Tx, T G R^^^ of order r exists if there exist f G T?''^", Ti G /?"><^ T2 G /?"><^ and ^ G i?^><^ such that for A stable. TA - AT - /TC
and
TiC + 727 = 7.
(3.11)
We thus have B = TB — KD, We note that it can be shown that for (A, C) detectable and r sufficiently large, there will always exist a solution. An example is the case when 7 = / a n d r = ^ - p . This is the case of the reduced-order observer discussed next, in Subsection B. Another simple case of interest is when 7 is a row vector {n = 1). In this case it can be shown (Luenberger [14]) that an observer of order i^ - 1 [i^ is the observability index of (A, C)] can always be constructed with its eigenvalues freely assignable, which will asymptotically estimate the linear function of the state, Tx. In general it can be shown that an observer of order h{v - 1) can be constructed that will asymptotically estimate Tx. Note that the problem of determining an observer of Tx of minimum order is a very difficult problem, except in special cases.
B. Reduced-Order Observers: Continuous-Time Systems Suppose thatp states, out of the n state, can be measured directly. This information can then be used to reduce the order of the full-state estimator from nton-p. Similar results are true for the estimator of a linear function of the state, but this problem will not be addressed here. To determine a full-state estimator of order n-p, first consider the case when C = [Ip, 0]. In particular, let Xi
=
z
"All A21
= [lp,0]
Ixi + A22. [X2_
An
Xil X2y
(3.12)
where z = xi represents the p measured states. Therefore, only X2(t) G R^ ^ is to be estimated. The system whose state is to be estimated is now given by X2 = A22X2 + [A2b B2] — ^22-^2 + Bu,
CHAPTER 4 :
State Feedback and State Observers
z - r i = Az + 5w + i^(Cx + Du) - r(A;c + 5w) =^ Az + (KC - tA)x + {B-fB
355
(3.13)
356 Linear Systems
where B == [A21, B2] and u =
is a known signal. Also,
^ A . (3.14) Anxi - Biu = A12X2, y = xi where y is known. An estimator for X2 can now be constructed. In particular, in view of (3.2), we have that the system X2 = A22X2 + Bu + K(y - A12X2) = (A22 - KAn)x2 + (A21Z + B2U) + m
- Anz - Biu)
(3.15)
is an asymptotic state estimator for X2. Note that the error e satisfies the equation ^ = i:2 - ^2 = (A22 - KAx2)e,
(3.16)
and if (A22, A12) is observable, then the eigenvalues of A22 - ^Ai2 can be arbitrarily assigned making use of K. It can be shown that if the pair (A = [Aij\ C = [Ip, 0]) is observable, then (A22. A12) is also observable (prove this using the eigenvalue observability test of Section 3.4 of Chapter 3). System (3.15) is an estimator of order z . To avoid using z = xi n- p, and therefore the estimate of the entire state x is X2 in y given by (3.14), one could use X2 = w + Kz and obtain from (3.15) an estimator in terms of w, z, and u. In particular, w = (A22 - KAn)w + [(A22 - KAn)K
+ A21 - KAix\z + [B2 - KBi]u.
(3.17)
Then w is an estimate of X2 - Kz, and of course w + ^ z is an estimate for X2 (verify this). In the above derivation, it was assumed for simplicity that a part of the state xi, is measured directly, i.e., C = [Ip, 0]. One could also derive a reduced-order estimator for the system X = Ax + Bu, y = Cx.
\c To see this, let rank C = p and define a similarity transformation matrix P = \^ where C is such that P is nonsingular. Then k = Ax + Bu,y = Cx = [Ip, 0]x,
(3.18)
where x = Px,A = PAP'K B = PB, and C = CP'^ = [Ip, 0] (show this). The transformed system is now in an appropriate form for an estimator of order n p to be derived, using the procedure discussed above. The estimate of x is and the estimate of the original state x is P
. In particular, X2 = w + Ky,
where w satisfies (3.17) with z = y, [Atj] = A = PAP ^ and
Bi B2
B = PB.
The interested reader should verify this result. EXAMPLE 3.3. Consider the system x
ro
0 Ax + Bu, y = Cx, where A = 1
-2 -2
B = [i. , and C = [0, 1]. We wish to design a reduced n- p = n- I = 2- I = firstorder asymptotic state estimator.
C C. "0 1 0 - 2 "0 1 X = PxdindA = PAP~^ = 1 0 1 -2 1 0 C = CP ^ = [1,0]. The system {A, B, C} is now in All An - 2 1 ,B (3.17). We have A [A21 A22, -2 0 the form The similarity transformation matrix P
"0 1 leads to (3.18), where .1 0 -2 1 B = PB = and -2 0 an appropriate form for use of L and (3.17) assumes
w = (-K)w + [-K^ + (-2) - K(-2)]y + (-K)u, a system observer of order 1. For K = -10 we have w = lOw - I22y + lOw, and w + Ky = w - lOy is an estimate for X2. Therefore,
y is an estimate of x, and w - lOy
y [w — lOy
0 1 y 1 0 w - lOy
w - lOy y
is an estimate of x(t) for the original system. C. Optimal State Estimation: Continuous-Time Systems The gain K in the estimator (3.2) above can be determined so that it is optimal in an appropriate sense. This is discussed very briefly in the following. The interested reader should consult the extensive literature on filtering theory for additional information, in particular, the literature on the Kalman-Bucy filter. In addressing optimal state estimation, noise with certain statistical properties is introduced in the model and an appropriate cost functional is set up that is then minimized. In the following, we shall introduce some of the key equations of the Kalman-Bucy filter and we will point out the duality between the optimal control and estimation problems. We concentrate on the time-invariant case, although, as in the LQR control problem discussed earlier, more general results for the time-varying case do exist. We consider the linear time-invariant system X = Ax + Bu + Tw, y = Cx + V,
(3.19)
where w and v represent process and measurement noise terms. Both w and v are assumed to be white, zero-mean Gaussian stochastic processes, i.e., they are uncorrelated in time and have expected values £[w] = 0 and £[v] = 0. Let E[ww^] = W,
E[vv^] = V
(3.20)
denote their covariances, where W and V are real, symmetric, and positive definite matrices, i.e., W = W^, W > 0, and V = V^, V > 0. Assume that the noise processes w and V are independent, i.e., E[wv^] = 0. Also assume that the initial state ;c(0) of the plant is a Gaussian random variable of known mean, E[x(0)] = XQ, and known covariance, E[(x(0) - jco)(-^(0) ~ -^o)^] = PeO- Assume also that x(0) is independent of w and v. Note that all these are typical assumptions made in practice. Consider now the estimator (3.2), namely, X = Ax + Bu + K(y - Cx) = (A - KC)x + Bu + Ky,
(3.21)
357 CHAPTER 4 :
State Feedback and State Observers
358 Linear Systems
and let (A, F W^^^, C) be controllable (-from-the-origin) and observable. It turns out that the error covariance E[(x - x){x - x)^] is minimized when the filter gain is given by K* = PlC^V-\
(3.22)
where P* denotes the symmetric, positive definite solution of the quadratic (dual) algebraic Riccati equation P,A^ + APe - PeC^V'^CP,
+ TWT^
0.
(3.23)
Note that P*, which is in fact the minimum error covariance, is the positive semidefinite solution of the above Riccati equation if (A, TW^'^, C) is stabilizable and detectable. The optimal estimator is asymptotically stable. The above algebraic Riccati equation is the dual to the Riccati equation given in (2.34) for optimal control and can be obtained from (2.34) making use of the substitutions A^,B
C^,M
and
R
V,Q^W.
(3.24)
Clearly, methods that are analogous to the ones developed by solving the control Riccati equation (2.34) may be applied to solve the Riccati equation (3.21) in filtering. These methods are not discussed here. EXAMPLE 3.4 Consider the system i = Ax,y = CJC, whereA = and let T =
ro 01 1 0
X = [0, 1],
,V = p > 0,W = 1. We wish to derive the optimal filter K*
P*C'^V~^ given in (3.22). In this case, the Riccati equation (3.23) is precisely the Riccati equation of the control problem given in Example 2.7. The solution of this equation was determined to be P: We note that this was expected, since our example was chosen to satisfy (3.24). Therefore,
Jp
1
r = P: i j p
J^p
D. Full-Order Observers: Discrete-Time Systems We consider described by equations of the form x{k + 1) = Ax{k) + Bu{k),
y = Cx{k) + Du(k),
(3.25)
where A G 7^'^><^ B G i?^><'^, C G RP"""^, and D G RP"""^. The construction of state estimators for discrete-time systems is mostly analogous to the continuous-time case, and the results that we established above for such systems are valid in here as well, subject to obvious adjustments and modifications. There are, however, some notable differences. For example, in discrete-time systems it is possible to construct a state estimator that converges to the true value of the state in finite time, instead of infinite time as in the case of asymptotic state estimators.
This is the estimator known as the deadbeat observer. Furthermore, in discrete-time systems it is possible to talk about current state estimators, in addition to prediction state estimators. In what follows, a brief description of the results that are analogous to the continuous-time case are given. Current estimators and deadbeat observers that are unique to the discrete-time case are discussed at greater length. Full-state observers: The identity observer As in the continuous-time case, following (3.2) we consider systems described by equations of the form x{k + 1) = Ax{k) + Bu{k) + K[y{k) - y{k)l
(3.26)
where y{k) = Cx(k) -h Dx{k). This can also be written as x{k + 1) = (A - KC)x(k) + [5 - KD, K]
u(k)
(3.27)
It can be shown that the error e(k) = x(k) - x(k) obeys the equation e(k + 1) == (A - KC)e(k). Therefore, if the eigenvalues of A - KC are inside the open unit disc of the complex plane, then e(k) ^ 0 as /: —> oo. There exists K so that the eigenvalues of A - KC can be arbitrarily assigned if and only if the pair (A, C) is observable (see Lemma 3.1). The discussion following Lemma 3.1 for the case when (A, C) is not completely observable, although detectable, is still valid. Also, the remarks on appropriate locations for the eigenvalues of A - KC and noise being the limiting factor in state estimators are also valid in the present case. Note that the latter point should seriously be considered when deciding whether or not to use the deadbeat observer described next. To balance the trade-offs between speed of the estimator response and noise amplification, one may formulate an optimal estimation problem as was done in the continuous-time case, the Linear Quadratic Gaussian (LQG) design being a common formulation. The Kalman filter (discrete-time case) which is based on the "current estimator" described below is such a quadratic estimator. The LQG optimal estimation problem can be seen to be the dual of the quadratic optimal control problem discussed in the previous section. As in the continuous-time case, optimal estimation in the discrete-time case will be discussed only briefly in the following. First, however, several other related issues are addressed. Deadbeat observer. If the pair (A, C) is observable, it is possible to select K so that all the eigenvalues of A - KC are at the origin. In this case e{k) = x(k)- x{k) = (A - KC)^e(0) = 0, for some k^ n\ i.e., the error will be identically zero within at most n steps. The minimum value of k for which (A - KC)^ = 0 depends on the size of the largest block on the diagonal of the Jordan canonical form of A - KC. (Refer to the discussion on the modes of discrete-time systems in Section 2.7 of Chapter 2.) EXAMPLE 3.5. Consider the system x(^ + 1) = Ax(k), y(k) =
A=
C
p 1 0 1
CJC(^),
where
359 CHAPTER 4 :
State Feedback and State Observers
360 Linear Systems
is in observer form. We wish to design a deadbeat observer. It is rather easy to show (compare with Example 2.5) that " K
' 2 11"
Aj - - 1 0 r - 1 am
0 Oj.
0" [ 1 -1
which was determined by taking the dual AD = A^,BD = C^ in controller form, using FD = B;,'[Ad^ - Am] and K = -F^ The matrix Aj consists of the second and third columns of a matrix A^ = 0
X
1
X
0
X
in observer (companion) form with all its eigenvalues at 0. For Aj^
"0 0 0" 1 0 0 , we have .0 1 0_
ro 0' " 2 1 Ki = 1p 0 - - 1 0 0 0 Li 0.
-1 1
0 -1
1 1 -1 0 -1 0
, we obtain
and for A^2 ^
1 -1 0
K2
Note that A - /^iC = A^ A 2
_
and Al
0, and A - K2C = A^^, and
0. Therefore, for the observer gain Ki the error e(k) in the deadbeat observer will become zero in n = 3 steps since e(3) = (A - KiC)^e(0) = 0. For the observer gain K2, the error e(k) in the deadbeat observer will become zero in2 < n steps, since e(2) = (A - K2C)^e(0) = 0. The reader should determine the Jordan canonical forms of Ad^ and A^^ ^^^ verify that the dimension of the largest block on the diagonal is 3 and 2, respectively. •
Ai
The comments in the discussion following Lemma 3.1 on taking ^ = 0 are valid in the discrete-time case as well. Also, the approach of determining the state instantaneously in the continuous-time case, using the derivatives of the input and output, corresponds in the discrete-time case to determining the state from current and future input and output values (see Exercise 3.14 in Chapter 3). This approach was in fact used to determine x(0) when studying observability in Section 3.3. The disadvantage of this method is that it requires future measurements to calculate the current state. This issue of using future or past measurements to determine the current state is elaborated upon next. Current estimator The estimator (3.26) is called 3. prediction estimator. The state estimate x(k) is based on measurements up to and including y(/: - 1). It is often of interest in
applications to determine the state estimate x(k) based on measurements up to and including y(k). This may seem rather odd at first; however, if the computation time required to calculate x(k) is short compared to the sample period in a sampled-data system, then it is certainly possible practically to determine the estimate x(k) before x(k + 1) and y(k + 1) are generated by the system. If this state estimate, which is based on current measurements of y(k), is to be used to control the system, then the unavoidable computational delays should be taken into consideration. Now let x(k) denote the current state estimate based on measurements up through y(k). Consider the current estimator
where
x(k) = x(k) + Kc(y(k) - Cx{k)),
(3.28)
x(k) = Ax(k - 1) + Bu(k - 1),
(3.29)
i.e., x(k) denotes the estimate based on model prediction from the previous time estimate, x(k - 1). Note that in (3.28), the error is y(k) - y(k), where y(k) = Cx(k) (D = 0), for simplicity. Combining the above, we obtain x(k) = (I - KcC)Ax(k - 1) + [(/ - KcC)B, -Kc]
\u(k - 1)1 y(k)
(3.30)
The relation to the prediction estimator (3.26) can be seen by substituting (3.28) into (3.29) to obtain x(k + 1) = Ax(k) + Bu(k) + AKc[y(k) - Cx(k)l
(3.31)
Comparison with the prediction estimator (3.26) (with D = 0) shows that if K - AKc,
(3.32)
then (3.31) is indeed the prediction estimator, and the estimate x(k) used in the current estimator (3.28) is indeed the prediction state estimate. In view of this, we expect to obtain for the error e(k) = x(k) - x(k) the difference equation e(k + 1) = (A - AKcC)e(k)
(3.33)
(show this). To determine the error e{k) = x{k) - x(k) we note that e{k) = e(k) - (x(k) - x(k)). Equation (3.28) now implies that x(k) - x(k) = KcCe(k). Therefore, e(k) =
(I-KcC)e(kl
(3.34)
This establishes the relationship between errors in current and prediction estimators. Premultiplying (3.31) by / - KcC (assuming |/ - KcC\ ¥- 0), we obtain e(k^
1) -
{A-KcCA)e(k\
(3.35)
which is the current estimator error equation. The gain Kc is chosen so that the eigenvalues of A - KcC A are within the open unit disc of the complex plane. The pair (A, CA) must be observable for arbitrary eigenvalue assignment. Note that the two error equations (3.33) and (3.35) have identical eigenvalues (show this). EXAMPLE 3.6. Consider the system x{k + 1) = Ax(^), y{]i) = Cx(k), where A = 0 -21 C = [0, 1], which is in observer form (see also Example 3.2). We wish to 1
361 CHAPTER 4:
state Feedback and State Observers
362 Linear Systems
design a current estimator. In view of the error equation (3.35), we consider l\s 0" _ / "0 -2" po [1 - 2 ] | o s_ .1 - 2 . Ui. s + ko 2 - Iko ] = det ki - 1 s + 2-2y^iJ = / + 5(2 - 2y^i + ko) + (2 - 2y^i) = 5^ + dis -\- do = ad(s),
det{sl - (A - KcCA)) = det
I
a desired polynomial, from which Kc = [ko, ki]^ = [di - do, \{2 - Jo)]^- Note that AKc = [do -2,di - 2f = K, found in Example 3.2, as noted in (3.32). The current estimator (3.30) is now given by x{k) = {A - KcCA)x{k - 1) KcCBu(k - 1) + Kcy(kl or — ko — 2 + 2ko x(k) x(k - 1) + y(k). I - ki -2 + 2ki_ Partial or linear functional state observers The problem of estimating a linear function of the state, Tx{k), T E R^^^^ where ft ^ n, using a prediction estimator, is completely analogous to the continuous-time case and w^ill therefore not be discussed further here. E. Reduced-Order Observers: Discrete-Time Systems It is possible to estimate the full state x(k) using an estimator of order n- p, where p = rank C. If a prediction estimator is used for that part of the state that needs to be estimated, then the problem in the discrete-time case is completely analogous to the continuous-time case, discussed before. We will omit the details. F. Optimal State Estimation: Discrete-Time Systems The formulation of the Kalman filtering problem in discrete-time is analogous to the continuous-time case. Consider the linear time-invariant system given by x{k + 1) = Ax{k) + Bu{k) + Tw{k\
y(k) = Cx(k) + v,
(3.36)
where the process and measurement noises w, v are white, zero-mean Gaussian stochastic processes, i.e., they are uncorrected in time with £'[vv] = 0, andjE'[v] == 0. Let the covariances be given by E[ww^] = W,
E[vv^] = V,
(3.37)
where W = W^, W > 0 and V = V^, V > 0. Assume that w, v are independent, that the initial state x{0) is Gaussian of known mean {E[x{0)\ = XQ), that E[(x(0) — xo)(x(0) - XQ)^] = Peo, and that x(0) is independent ofw and v. Consider now the current estimator (3.26), namely, x(k) = x(k) + Kc[y(k) -
Cx(k)l
where x(k) = Ax(k - 1) + Bu(k - 1), and x(k) denotes the prior estimate of the state at the time of a measurement.
It turns out that the state error covariance is minimized when the filter gain is Kl = PlC^(CPlC^
+ V)~\
(3.38)
where P* is the unique, symmetric, positive definite solution of the Riccati equation Pe -
A[Pe
- PeC^iCPeC^
+ W^CPeU^
+ TWT^.
(3.39)
It is assumed here that (A, FW^^^, C) is reachable and observable. This algebraic Riccati equation is the dual to the Riccati equation (2.53) that arose in the discretetime LQR problem and can be obtained by substituting A^,B
C^,M
and
R-^
V,Q
W.
(3.40)
It is clear that, as in the case of the LQR problem, the solution of the algebraic Riccati equation can be determined using the eigenvectors of the (dual) Hamiltonian. The filter derived above is called the discrete-time Kalman filter. It is based on the current estimator (3.28). Note that AKc yields the gain K of the prediction estimator [see (3.32)]. 4.4 OBSERVER-BASED DYNAMIC CONTROLLERS State estimates, derived by the methods described in the previous section, may be used in state feedback control laws to compensate given systems. This section addresses that topic. In Section 4.2, the linear state feedback control law was introduced. There it was implicitly assumed that the state vector x{t) is available for measurement. The values of the states x{t) for t > t^ were fed back and used to generate a control input in accordance with the relation u{t) = Fx(t) + r(t). There are cases, however, when it may be either impossible or impractical to measure the states directly. This has provided the motivation to develop methods for estimating the states. Some of these methods were considered in Section 4.3. A natural question that arises at this time is the following: what would happen to system performance if, in the control law u = Fx-\-r, the state estimate x were used in place of x as in Fig. 4.5? How much, if any, would the compensated system response deteriorate? What are the difficulties in designing such estimator- (observer-) based linear state feedback controllers? These questions are addressed in this section. Note that observer-based controllers of the type described in the following are widely used.
' Itr v^* 1
•
^ u
1
-{-
*
System [
1
State observer F 1FIGURE 4.5 Observer- 3ased contro Her
U
363 CHAPTER 4 :
State Feedback and State Observers
364 Linear Systems
In the remainder of this section we will concentrate primarily on full-state/ full-order observers and (static) linear state feedback, as applied to linear timeinvariant systems. The analysis of partial-state and/or reduced-order observers with static or dynamic state feedback is analogous; however, it is more complex. Such control schemes are not considered at this point. Instead, they will be discussed in the exercise section (Exercise 4.8). In this section, continuous-time systems are addressed. The analysis of observer-based output controllers in the discrete-time case is completely analogous and will be omitted.
A. State-Space Analysis We consider systems described by equations of the form x = Ax-\-Bu,
y = Cx + Du,
(4.1)
where A E i?"^^'^, B G T^^^^^, C G i^^^^^, and D G RP'^'^. For such systems, we determine an estimate x{t) G R^ of the state x{t) via the (full-state/full-order) state observer (3.2) given by = Ax^Bu-\= (A=
K(y - y)
KC)x + [J5 - KD, K] u (4.2)
X,
where y = Cx + Du. We now compensate the system by state feedback using the control law u = Fx + r,
(4.3)
where x is the output of the state estimator and we wish to analyze the behavior of the compensated system. To this end we first eliminate y in (4.2) to obtain X = {A- KC)x + KCx + Bu,
(4.4)
The state equations of the compensated system are then given by X = Ax-^ BFx + Br k = KCx + (A-KC
+ BF)x + Br,
(4.5)
and the output equation assumes the form y = Cx + DFx + Dr,
(4.6)
where u was eliminated from (4.1) and (4.4), using (4.3). Rewriting in matrix form, we have X X
=
A KC
BF 1 \x + B B KC + BF\ IX
A
y = [QDF]
X X
+ Dr,
(4.7)
which is a representation of the compensated closed-loop system. Note that (4.7) constitutes a 2Azth-order system. Its properties are more easily studied if an appropriate
similarity transformation is used to simplify the representation. Such a transformation is given by X
PKX
==
7
01 \x
/
-l\ [x
=
X
e
(4.8)
where the error e(t) = xit) - x(t). Then the equivalent representation is X
e
=
A + BF 0
B -BF \x + 0 A-KC [e
y = [C + DF, -DF]
x\
+ Dr,
(4.9)
It is now quite clear that the closed-loop system is not fully controllable with respect to r (explain this in view of Subsection 3.4A). In fact, e{t) does not depend on r at all. This is of course as it should be, since the error e{t) = x(t) - x(t) should converge to zero independently of the externally applied input r. The closed-loop eigenvalues are the roots of the polynomial \sln - (A + BF)\\sIn - (A - KC)\.
(4.10)
Recall that the roots of \sln - (A + BF)\ are the eigenvalues of A + BF that can arbitrarily be assigned via F provided that the pair (A, B) is controllable. These are in fact the closed-loop eigenvalues of the system when the state x is available and the linear state feedback control law u = Fx-\- ris used (see Section 4.2). The roots of |^/„ - (A - KC)\ are the eigenvalues of (A - KC) that can arbitrarily be assigned via K provided that the pair (A, C) is observable. These are the eigenvalues of the estimator (4.2). The above discussion points out that the design of the control law (4.3) can be carried out independently of the design of the estimator (4.2). This is referred to as the Separation Property and is generally not true for more complex systems. The separation property indicates that the linear state feedback control law may be designed as if the state x were available and the eigenvalues of A + BF are assigned at appropriate locations. The feedback matrix F can also be determined by solving an optimal control problem (LQR). If state measurements are not available for feedback, a state estimator is employed. The eigenvalues of a full-state/full-order estimator are given by the eigenvalues of A - KC. These are typically assigned so that the error e(t) = x(t) - x(t) becomes adequately small in a short period of time. For this, the eigenvalues of A - KC are (empirically) taken to be about 6 to 10 times further away from the imaginary axis (in the complex plane, for continuous-time systems) than the eigenvalues of A + BF. The behavior of the closed-loop system should be verified since the above is only a rule of thumb. (Refer to any good book on control for further discussion—see Section 4.6, Notes.) The estimator gain K may also be determined by solving an optimal estimation problem (the Kalman filter). In fact, under the assumption of Gaussian noise and initial conditions given earlier (see Section 4.3), F and i^ can be found by solving, respectively, optimal control and estimation problems with quadratic performance criteria. In particular, the deterministic LQR problem is first solved to determine the optimal control gain F*, and then the stochastic Kalman filtering problem is solved to determine the optimal filter gain i^*. The separation property (i.e.. Separation Theorem—see any optimal control textbook) guarantees that the overall (state estimate feedback) Linear Quadratic Gaussian (LQG)
365 CHAPTER 4 :
State Feedback and State Observers
366 Linear Systems
control design is optimal in the sense that the control law w*(0 = F''x{t) minimizes the quadratic performance index E[\^{z^Qz + u^Ru) dt]. As was discussed in previous sections, the gain matrices F* and K* are evaluated in the following manner. Consider X = Ax-\- Bu-l-Tw,
y = Cx-\-v,z
= Mx
(4.11)
with E[ww^] = W > 0 and E[vv'^] = V > 0 and with Q > 0, R > 0 denoting the matrix weights in the performance index E[\^(z^Qx + u^Ru) dt]. Assume that both (A, B, Q^'^M) and (A, TW^''^, C) are controllable and observable. Then the optimal control law is given by u{t) = F'^m
= -R-^B^Pim,
(4.12)
where P* > 0 is the solution of the algebraic Riccati equation (2.34) given by A^Pc + PcA - PcBR-^B^Pc
+ M^QM = 0.
(4.13)
The estimate x is generated by the optimal estimator k = Ax^
Bu + K\y
- Cx),
(4.14)
^* = Plc^y-
where
(4.15)
in which P* > 0 is the solution to the dual algebraic Riccati equation (3.21) given by PeA^
+ APe
- PeC^V'^CPe
+ TWT^
= 0.
(4.16)
Designing observer-based dynamic controllers by the LQG control design method has been quite successful, especially when the plant model is accurately known. In this approach the weight matrices Q, R and the covariance matrices W, V are used as design parameters. Unfortunately, this method does not necessarily lead to robust designs when uncertainties are present. This has led to an enhancement of this method, called the LQR/LTR (Loop Transfer Recovery) method, where the design parameters W and V are selected (iteratively) so that the robustness properties of the LQR design are recovered (refer to Section 4.6). Finally, as was mentioned, the discrete-time case is analogous to the continuoustime case and its discussion will be omitted. EXAMPLE 4.1. Consider the system x = Ax + Bu, y = Cx, where 0 1 0 ro 11 1 1 C = [1, 0, 0]. 0 0 1 .0 0. 0 2 - 1 This is a controllable and observable but unstable system with eigenvalues of A equal to 0, - 2 , 1. A linear state feedback control u = Fx -\- r was derived in Example 2.5 to assign the eigenvalues of A + BF at - 2 , - 1 ± j . An appropriate F to accomplish this was shown to be F=
2 -2
-1 0
-2 \
If the state x(t) is not available for measurement, then an estimate x(t) is used instead, i.e., the control law u = Fx + r is employed. In Example 3.1, a full-order/full-state
observer, given by
367 X = (A-
CHAPTER 4 :
KC)x + [B, K^
was derived [see (3.3)] with the eigenvalues of A - KC determined as the roots of the polynomial ad{s) = s^ + d2S^ + dis + do. It was shown that the (unique) K is in this case K = [d2~ 1, Ji - ^2 + 3, do-di+
3d2 - 5]^,
and the observer is given by 0 1 d2-l 1 0" 0 1 x+ 1 1 ^1-^2+3 2 -1. 0 0 do-di+ 3d2 - 5
1 -d2 -di +d2-3 _-do + di - 3^2 + 5
Using the estimate x in place of the control state x in the feedback control law causes some deterioration in the behavior of the system. This deterioration can be studied experimentally. (See the next subsection for analytical results.) To this end, let the eigenvalues of the observer be at, say, - 1 0 , - 1 0 , - 1 0 , let x(0) = [1, 1, 1]^ and x(0) = [0, 0, 0]^, plot x(t), x(t), and e(t) = x(t) — x(t), and compare these with the corresponding plots of Example 2.5, where no observer was used. Repeat the above with observer eigenvalues closer to the eigenvalues of A + BF (say, at - 2 , - 1 ± j) and also further away. In general the faster the observer, the faster e(t) -^ 0, and the smaller the deterioration of response; however, in this case care should be taken if noise is present in the system. •
B. Transfer Function Analysis For the compensated system (4.9) [or (4.7)], the closed-loop transfer function T(s) between y and r is given by 3;(^) = T(s)r(s)
= [(C + DF)[sI
- (A + BF)Y^B
+ D]r{s),
(4.17)
where y{s) and r{s) denote the Laplace transforms of y{t) and r ( 0 , respectively. The function T{s) was found from (4.9), using the fact that the uncontrollable part of the system does not appear in the transfer function (see Section 3.4). Note that T{s) is the transfer function of {A + BF, B,C + DF, D], i.e., T{s) is precisely the transfer function of the closed-loop system Hp{s) when no state estimation is present (see Section 4.2). Therefore, the compensated system behaves to the outside world as if there were no estimator present. Note that this statement is true only after sufficient time has elapsed from the initial time, allowing the transients to become negligible. (Recall what the transfer function represents in a system.) Specifically, taking Laplace transforms in (4.9) and solving, we obtain
=
[si -(A
+ BF)] - 1 0
_ -[si
+ [si -(A + BF)]-^B 0
(A + BF)]-^BF[sI - (A - KC)]-^] [si - (A-h BF)]-^
\x(0) \\e(0)
ris)
Ks) = [C + DF, -DF] Ms)] + Dris). _e(s)\
(4.18)
State Feedback and State Observers
368
Therefore,
Linear Systems
y(s) - ( C + DF)[sI - (A +
BF)r^x(0)
- [(C + DF)[sI - (A + BF)r^BF[sI + DF[sI - (A + BF)r^]e(0)
- (A
KC)y
+ T(s)r{s\
(4.19)
which indicates the effects of the estimator on the input-output behavior of the closed-loop system. Notice how the initial conditions for the error ^(0) = x(0) - x(0) influence the response. Specifically, when ^(0) T^ 0, its effect can be viewed as a disturbance that will become negligible at steady state. The speed by which the effect of ^(0) on y will diminish depends on the location of the eigenvalues of A + BF and A - KC, as can be easily seen from relation (4.19). Two-input controller In the following, we will find it of interest to view the observer-based controller discussed previously as a one-vector output (u) and a two-vector input (y and r) controller. In particular, from x = (A - KC)x -\- (B - KD)u + Ky given in (4.2) and u = Fx -\- r given in (4.3), we obtain the equations X = {A-KC
+ BF - KDF)x + [K,B-
KD] (4.20)
u = Fx + r.
This is the description of the (nth order) controller shown in Fig. 4.5. The state x is of course the state of the estimator and the transfer function between u and y, r is given by u(s) = F[sl -{A-KC + [F[sl -{A-
+ BF -
KDF)Y^Ky{s)
KC + BF ~ KDF)Y\B
- KD) 4- I]r{s\
(4.21)
If we are interested only in "loop properties," then r can be taken to be zero, in which case (4.21) (for r = 0) yields the output feedback compensator, which accomplishes the same control objectives (that are typically only "loop properties") as the original observer-based controller. This fact is used in the LQG/LTR design approach. When r 7^ 0, (4.21) is not appropriate for the realization of the controller since the transfer function from r, which must be outside the loop, may be unstable. Note that an expression for this controller that leads to a realization of a stable closed-loop system is given by u{s) = [F[sl -{A-KC
+ BF - KDF)Y^[K, B - KD] -I- [0, /]]
r(s)
(4.22)
(see Fig. 4.6). This was also derived from (4.20). The stability of general two-input controllers (with two degrees of freedom) is discussed at length in Section 7.4D of Chapter 7.
r
Controller
u
System
y
FIGURE 4.6 Two-input controller
At this point, we find it of interest to determine the relationship of the observerbased controller and the conventional one-and-two block controller configurations of Fig. 4.7. Here, the requirement is to maintain the same transfer functions between inputs y and r and output u. (For further discussion of stability and attainable response maps in systems controlled by output feedback controllers, refer to Section 7.4 of Chapter 7.) We proceed by considering once more (4.2) and (4.3) and by writing u(s) = F[sl - (A - KC)r\B
-
KD)u(s)
+ F[sl - (A - KC)r^Ky(s)
+ r(s) = Guu(s) + Gyy(s) + r(s).
This yields u(s) = (I-
Gur'[Gyy(s)
+ r(s)]
(4.23)
(see Fig. 4.7). Notice that Gy = F[sl - (A - KOr^K,
(4.24)
i.e., the controller in the feedback path is stable (why?). The matrix (/ - Gu)~^ is not necessarily stable; however, it is inside the loop and the internal stability of the compensated system is preserved. Comparing with (4.21), we obtain (/ - G J ~ i = F[sl -{A-KC
+ BF - KDF)Y\B
- KD) + /.
(4.25)
Also, as expected, we have (/ - GuT^Gy = F[sl -{A-KC^-BF
- KDF)Y^K
(4.26)
(show this). These relations could have been derived directly as well by the use of matrix identities; however, such derivation is quite involved. r
' '^0 • 1 + 1 1 1
(1 - Gu)-'
'
u
System
y
Gy
FIGURE 4.7 EXAMPLE 4.2. For the system x = Ax + Bu, y = Cx with A =
B=
and C = [0, 1] we have H(s) = C(sl - A)-^B = s/(s^ + 2s + 2). In Example 3.2, it was shown that the gain matrix K = [do - 2, di - 2]^ assigns the eigenvalues of the asymptotic observer (of A - KC) at the roots of s^ + dis + do. In fact si - (A- KC) = s do . It is straightforward to show that F = [\ao — 1,2 — ai\ will assign the -1 s + di_ eigenvalues of the closed-loop system (of A + BF) at the roots of ^^ + ^i^ + ao. Indeed, s 2 si -{A + BF) . Now in (4.23) we have kao s -\- a\ Gy{s) = F(sI-(A-KC))-^K s((do - 2)(iao - 1) + (di - 2)(2 - aQ) + ((do - d,){ao - 2) + {do - 2)(2 - a,)) s^ + dis + do
369 CHAPTER 4 :
State Feedback and State Observers
370
1
s(2 — a\) — doilao — I) = ^ ^'' °^^ ° '-, s"-+ dis + do (1 - G„)-' = ''+d,s + do _ s^ + s(di + a\ - 2) + haodo G,(s) = FisI - (A - KOr'B
Linear Systems
4.5 SUMMARY In this chapter, Hnear state feedback controllers and state observers were studied with an emphasis on time-invariant, continuous-time, and discrete-time systems (Sections 4.2 and 4.3). State feedback controllers and state observers were then combined (in Section 4.4) to develop observer-based dynamic output feedback controllers. We note that output feedback controllers are studied further in Chapter 7. Linear state feedback was studied in Section 4.2. First, the need for feedback was explained by discussing open- and closed-loop control in the presence of uncertainties. System stabilization was then addressed which for the time-invariant case leads to the eigenvalue or pole assignment problem. It was pointed out that an optimal control formulation, such as the Linear Quadratic Regulator (LQR) problem, will also lead to stable closed-loop control systems while attaining additional control goals as well. Furthermore, the LQR problem formulation can also be used in the time-varying case. The eigenvalue assignment problem was studied at length by introducing several such methods. Analogous results for the discrete-time case were established in Subsections 4.2E and 4.2F. In Section 4.3, full-order observers for the entire state or for a linear function of the state were presented. Reduced-order observers and optimal observers were also addressed in Subsections 4.3B and 4.3C, respectively. The duality between the state feedback controller and state observer problems was explored and emphasized. Analogous results for the discrete-time case were also described. Current state estimators were introduced and developed in Subsection 4.3D. In Section 4.4, state observers, together with state feedback controllers were used to derive dynamic output feedback controllers and the Separation Principle was discussed. The degradation of performance in state feedback control when an observer is used to estimate the state was explained. An analysis of the closed-loop system was carried out, using both state-space and transfer function matrix descriptions.
4.6 NOTES The fact that if a system is (state) controllable, then all its eigenvalues can arbitrarily be assigned by means of linear state feedback has been known since the 1960s. Original sources include Rissanen [19], Popov [17], and Wonham [23]. (See also remarks in Kailath [10], pp. 187, 195.) The present approach for eigenvalue assignment via linear state feedback, using the controller form, follows the development in Wolovich [22]. Ackermann's formula first appeared in Ackermann [1]. The development of the eigenvector formulas for the feedback matrix that assign all the closed-loop eigenvalues and (in part) the corresponding eigenvectors follows
Moore [16]. The corresponding development that uses (A, B) in controller (companion) form and polynomial matrix descriptions follows Antsaklis [3]. Related results on static output feedback and on polynomial and rational matrix interpolation can be found in Antsaklis and Wolovich [4] and Antsaklis and Gao [5]. Note that the flexibility in assigning the eigenvalues via state feedback in the multi-input case can be used to assign the invariant polynomials of si - {A -\r BF)\ conditions for this are given by Rosenbrock [20]. The Linear Quadratic Regulator (LQR) problem and the Linear Quadratic Gaussian (LQG) problem have been studied extensively, particularly in the 1960s and early 1970s. Sources for these topics include the books by Anderson and Moore [2], Kwakemaak and Sivan [11], Lewis [12], and Dorato et al. [9]. Early optimal control sources include Athans and Falb [6], and Bryson and Ho [8]. A very powerful idea in optimal control is the Principle of Optimality, Bellman [7], which can be stated as follows: "An optimal trajectory has the property that at any intermediate point, no matter how it was reached, the remaining part of a trajectory must coincide with an optimal trajectory, computed from the intermediate point as the initial point". For historical remarks on this topic, refer, e.g., to Kailath [10], pp. 240-241. The most influential work on state observers is the work of Luenberger. Although the asymptotic observer presented here is generally attributed to him, Luenberger's Ph.D. thesis work in 1963 was closer to the reduced-order observer presented above. Original sources on state observers include Luenberger [13], [14], and [15]. For an extensive overview of observers, refer to the book by O'Reilly [18]. When linear quadratic optimal controllers and observers are combined in control design, a procedure called LQG/LTR (Loop Transfer Recovery) is used to enhance the robustness properties of the closed loop system. For a treatment of this procedure, see Stein and Athans [21] and contemporary textbooks on multivariable control.
4.7 REFERENCES 1. J. Ackermann, "Der Entwurf linearer Regelungssysteme im Zustandsraum," Regelungstechnik und Prozessdatenverarheitung, Vol. 7, pp. 297-300, 1972. 2. B. D. O. Anderson and J. B. Moore, Optimal Control Linear Quadratic Methods, Prentice-Hall, Englewood Cliffs, NJ, 1990. 3. R J. Antsaklis, "Some New Matrix Methods Applied to Multivariable System Analysis and Design," Ph.D. Dissertation, Brown University, May 1976. 4. P. J. Antsaklis and W. A. Wolovich, "Arbitrary Pole Placement Using Linear Output Feedback Compensation," Int. J. Control, Vol. 25, No. 6, pp. 915-925, 1977. 5. P. J. Antsaklis and Z. Gao, "Polynomial and Rational Matrix Interpolation: Theory and Control Applications," Int. J. of Control, Vol. 58, No. 2, 349-404, August 1993. 6. M. Athans and P. L. Falb, Optimal Control, McGraw-Hill, New York, 1966. 7. R. Bellman, Dynamic Programming, Princeton University Press, Princeton, NJ, 1957. 8. A. E. Bryson and Y. C. Ho, Applied Optimal Control, Hoisted Press, New York, 1968. 9. P. Dorato, C. Abdallah and V Cerone, Linear-Quadratic Control: An Introduction, Prentice-Hall, Englewood Cliffs, NJ, 1995. 10. T Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980. 11. H. Kwakemaak and R. Sivan, Linear Optimal Control Systems, Wiley, New York, 1972. 12. F L. Lewis, Optimal Control, Wiley, New York, 1986.
371 CHAPTER 4 :
State Feedback and State Observers
372 Linear Systems
13. D. G. Luenberger, "Observing the State of a Linear System," IEEE Trans. Mil. Electron, MIL-8, pp. 74-80, 1964. 14 D. G. Luenberger, "Observers for Multivariable Systems," IEEE Trans, on Auto. Control, A C - l l , p p . 190-199, 1966. 15 D. G. Luenberger, "An Introduction to Observers," IEEE Trans, on Auto. Control, AC-16, pp. 596-603, December 1971. 16 B. C. Moore, "On the Flexibihty Offered by State Feedback in Multivariable Systems Beyond Closed Loop Eigenvalue Assignment," IEEE Trans, on Auto. Control, AC-21, pp. 689-692, 1976; see also AC-22, pp. 140-141, 1977 for the repeated eigenvalue case. 17 V. M. Popov, "Hyperstability and Optimality of Automatic Systems with Several Control Functions," Rev. Roum. Sci. Tech. Sen ElectrotechEnerg., Vol. 9, pp. 629-690,1964. See also V. M. Popov, Hyperstability of Control Systems, Springer-Verlag, New York, 1973. 18 J. O'Reilly, Observers for Linear Systems, Academic Press, New York, 1983. 19. J. Rissanen, "Control System Synthesis by Analogue Computer Based on the Generalized Linear Feedback Concept," in Proc. of Symp. on Analog Comp. Applied to the Study of Chem. Processes, pp. 1-13, Intern. Seminar, Brussels, 1960. Presses Academiques Europeennes, Bruxelles, 1961. 20 H. H. Rosenbrock, State-Space and Multivariable Theory, Wiley, New York, 1970. 21 G. Stein and M. Athans, "The LQG/LTR Procedure for Multivariable Feedback Control Design," IEEE Trans, on Automatic Control, Vol. AC-32, pp. 105-114, February 1987. 22, W. A. Wolovich, Linear Multivariable Systems, Springer-Verlag, New York, 1974. 23, W. M. Wonham, "On Pole Assigment in Multi-Input Controllable Linear Systems," Vol. AC-12, pp. 660-665, 1967.
4.8 EXERCISES 4.1. Consider the system x = Ax + Bu, where A
-0.01 0
1 1 with u = Ex. 0.25 0.75J (a) Verify that the three different state feedback matrices given by -1.1 - 3 . 7 0 0 0 0 -1.1 1.2333
0 -0.02
-0.1 0
and B
0 -0.1
all assign the closed-loop eigenvalues at the same locations, namely, at —0.1025 ± jO.04944. Note that in the first control law (Fi) only the first input is used, while in the second law (F2) only the second input is used. For all three cases plot x(t) = [xi(t), X2(t)]^ when x(0) = [0, 1]^ and comment on your results. This example demonstrates how different the responses can be for different designs even though the eigenvalues of the compensated system are at the same locations, (b) Use the eigenvalue/eigenvector assignment method to characterize all F that assign the closed-loop eigenvalues at -0.1025 ± 70.04944. Show how to select the free parameters to obtain Fi, F2, and F3 above. What are the closed-loop eigenvectors in these cases? 4.2. For the system x = Ax -\- Bu with A E /^"><" and B E 7?"^"", where (A, B) is controllable and m > I, choose u = Fx dis the feedback control law. It is possible to assign all eigenvalues of A + BE by first reducing this problem to the case of eigenvalue assignment for single-input systems (m = 1). This is accomplished by first reducing the system to a single-input controllable system. We proceed as follows.
Let F = g ' f, where g G R^ and / ^ G i?" are vectors to be selected. Let g be chosen such that (A, 5^) is controllable. T h e n / in A + 5F = A + (5g)/ can be viewed as the state feedback gain vector for a single-input controllable system (A, Bg), and any of the single-input eigenvalue assignment methods can be used to select/ so that the closed-loop eigenvalues are at desired locations. The only question that remains to be addressed is whether there exists g such that (A, Bg) is controllable. It can be shown that if (A, B) is controllable and A is cyclic, then almost any g G R^ will make (A, Bg) controllable. (A matrix A is cyclic if and only if its characteristic and minimal polynomials are equal.) In the case when A is not cyclic, it can be shown that if (A, B, C) is controllable and observable, then for almost any real output feedback gain matrix //, A + BHC is cyclic. So initially, by an almost arbitrary choice of H or F = HC, the matrix A is made cyclic, and then by employing a g, (A, Bg) is made controllable. The state feedback vector g a i n / is then selected so that the eigenvalues are at desired locations. Note that F = gf is always a rank one matrix, and this restriction on F reduces the applicability of the method when requirements in addition to eigenvalue assignment are to be met. For the present approach, see W. M. Wonham, "On Pole Assignment in MultiInput Controllable Linear Systems," IEEE Trans. Autom. Control, Vol. AC-12, pp. 660-665, December 1967, and F. M. Brasch and J. B. Pearson, "Pole Placement Using Dynamic Compensators," IEEE Trans. Autom. Control, Vol. AC-15, pp. 34-43, February 1970. For a discussion of cyclicity of matrices see Chapter 2, and also P. J. Antsaklis, "Cyclicity and Controllability in Linear Time-Invariant Systems," IEEE Trans. Autom. Control, Vol. AC-23, pp. 745-746, August 1978. (a) For A, B as in Exercise 4.4, use the method described above to determine F so that the closed-loop eigenvalues are at —1 ± jand—2 ± /Comment on your choice for ^. 0 11 (b) For A characterize all g such that the closed-loop and B
,1 ij
eigenvalues are at - 1 . 4.3. Consider the system x(k + 1) A =
Ax(k) • Bu(k), where
4 -1 0
0 0 1
0 0" 1 0 . - 1 1. '
B =
Determine a linear state feedback control law u(k) = Fx{k) such that all the eigenvalues of A + BE are located at the origin. To accomplish this, use (a) reduction to a single-input controllable system, (b) the controller form of (A, B), (c) det (zl - (A + BF)) and the resulting nonlinear system of equations. In each case, plot x(k) with x(0) = [1, 1, 1]^ and comment on your results. In how many steps does your compensated system go to the zero state? 4.4. For the system x = Ax + Bu, where [O
1
0
'l 0 0 0
o" 0 0 1
determine F so that the eigenvalues of A + BF are at - -1 ± J and - 2 ± / Use as many different methods to choose F as you can.
373 CHAPTER 4 :
State Feedback and State Observers
374
4.5. Consider the system x = Ax + Bu, where 0 2 1 0 2 0
Linear Systems
0 1 0 0 -1 0
-6 0 6 0 0 -2
3 -1 -2 0 2 0
r 0
-1 -1 1 0 1 -1
B =
-1 0 0 1 0
r
-2 -1 0 2 0
(a) Show that - 1 is an uncontrollable eigenvalue. (b) For the control law u = Fx, determine F so that the controllable eigenvalues of A + BF are -.l,-.2,-l±j,-2. 4.6. Consider the SISO system Xc = AcXc + BcU,y = CcXc + DcU, where {Ac, Be) is in controller form with 0 1 ••• 0 1 [01 Ac =
Be =
0 l-ao
0 -ai
••• •••
Cc = [Co, Ci, . . ., Cn-ll
1 -an-ij
and let w = FcX + r = [/o, / i , . . . , fn-i]x + r be the linear state feedback control law. Use the Structure Theorem of Chapter 3 to show that the open-loop transfer function is n-l
H(s) = Cc{sI-Ac)-'Bc
+ Dc
-iss^ + an-is^
• • • + CiS + Co
^ + • • • + ais + ao
+ Dc
_ n{s)^ d(s) and the closed-loop transfer function is HF(S) = (Cc + DcFc)[sI - (Ac + BcFc)]-'Bc + Dc _
(Cn-l
+ Dcfn-l)s''-'
s^ + (a„-i
+ • • • + (CiS + Dcfx)s
+ (CQ + Z ) , / o )
/„_!>«-! + ••• + ( « ! - / i > + (ao - /o)
-hD,
^FW
Observe that state feedback does not change the numerator n(s) of the transfer function, but it can arbitrarily assign any desired (monic) denominator polynomial dF(s) = d(s)-Fells, . Thus, state feedback does not (directly) alter the zeros of H(s), but it can arbitrarily assign the poles of H(s). Note that these results generalize to the MIMO case [see (2.44)]. 4.7. Consider the system x = Ax + Bu,y = Cx, where ro 1 01 [01 A = \o 0 1 , B = C = [1, 2, 0]. 1 0 -ij (a) Determine an appropriate linear state feedback control law u = Fx + Gr(G E R) so that the closed-loop transfer function is equal to a given desired transfer function ^ - W = s^ + \s + 2We note that this is an example of model matching, i.e., compensating a given system so that it matches the input-output behavior of a desired model. In the present
case, state feedback is used; however, output feedback is more common in model matching. (b) Is the compensated system in (a) controllable? Is it observable? Explain your answers. (c) Repeat (a) and (b) by assuming that the state is not available for measurement. Design an appropriate state observer, if possible. 4.8. Consider the nth order system x = Ax + Bu, y = Cx -\- Du. (a) Let z = Az + Ky + [FB - KD]u, A E R^'^^, be an asymptotic estimator of the linear function of the state, Fx, where F G R^^"- is the desired state feedback gain. Consider the control law u = z + r and determine the representation and properties of the closed-loop system. (b) Consider a reduced-order observer of order n - p (as in Subsection 4.3B) and the control law u = Fx + r. Determine the representation and properties of the closedloop system. Hint: The analysis is analogous to the full-state observer case given in Section 4.4. 4.9. Design an observer for the oscillatory system i ( 0 = v{t),v{t) = -COQX(0, using measurements of the velocity v. Place both observer poles at 5* = -COQ. 4.10. Consider the undamped harmonic oscillator xi(t) = X2(t), X2(t) = -(olxi(t) + u(t). Using an observation of velocity y = X2, design an observer/state feedback compensator to control the position xi. Place the state feedback controller poles at ^ = —(Oo± j(x)o and both observer poles a.ts= -COQ. Plot x(t) for x(0) = [1, 1]-^ and COQ = 2. 4.11. A servomotor that drives a load is described by the equation (d^O/dt^) + (dOldt) = w, where 6 is the shaft position (output) and u is the applied voltage. Choose u so that 6 and (dO/dt) will go to zero exponentially (when their initial values are not zero). To accomplish this, proceed as follows. (a) Derive a state-space representation of the servomotor. (b) Determine linear state feedback, u = Fx + r, so that both closed-loop eigenvalues are at - 1 . Such F is actually optimal since it minimizes J = IQW^ + (dO/dtf + u^] dt (show this). (c) Since only 9 and u are available for measurement, design an asymptotic state estimator (with eigenvalues at, say, - 3 ) and use the state estimate x in the linear state feedback control law. Write the transfer function and the state-space description of the overall system and comment on stability, controllability, and observability. (d) Plot 0 and dSldt in (b) and (c) for r = 0 and initial conditions equal to [1,1]^. (e) Repeat (c) and (d), using a reduced-order observer of order 1. 4.12. Consider the LQR problem for the system x = Ax + Bu, where (A, B) is controllable and the performance index is given by J{u) = \
Jo
e^'''[x^(t)Qx(t)-\-u^(t)Ru(t)]dt,
where a G /?, a > 0 and g > 0, /? > 0. (a) Show that u* that minimizes J(u) is a fixed control law with constant gains on the states, even though the weighting matrices Q = e^^^Q, R = e^^^R are time varying. Derive the algebraic Riccati matrix equation that characterizes this control law. (b) The performance index given above has been used to solve the question of relative stability. In the light of your solution, how do you explain this? Hint: Reformulate the problem in terms of the transformed variables x = e^^x.
375 CHAPTER 4 :
State Feedback and State Observers
376
4.13. Consider the system x --
Linear Systems
x+
u and the performance indices / i , J2 given by
/»oo
Ji=
/»oo
{x\+X2 + u^)dt
and
J2=
(900ix\+X2) + u^)dt. Jo Jo Determine the optimal control laws that minimize 7i and J2. In each case plot u{t), xi{t),X2{t) for x(0) = [1,1]^ and comment on your results. 4.14. Consider the discrete-time system x{k-\-l) =Ax{k) -\-Bu{k),y{k) = C{k)x{k), where 1x2 2^
C=[1,0]
and where T is the sampling period. This is a sampled-data system obtained from (the double integrator) A
.B
and C = [1,0] via a zero-order hold and an
ideal sampler, both of period T (see Chapter 2). Consider the cost functional J{u) = Yk=o(y^(^) + 2w^(^))- This represents (2.51) with z{k) = y{k),M = C,Q = 1, and R = 2. Determine the control sequence {u*{k)},k>0, that minimizes the cost functional, as a function of T. Compare your results with Examples 2.7 and 2.9. 0 1 u,y= [l,0]x. 1 0 x+ (a) Use state feedback u = Fx to assign"the eigenvalues of A+ BF at —0.5zbj0.5. Plot x(t) = [xi(t),X2{t)]'^ for the open- and closed-loop system with x(0) = [-0.6,0.4]^. (b) Design an identity observer with eigenvalues at — a ± y, where a > 0. What is the observer gain K in this case? (c) Use the state estimate x from (b) in the linear feedback control law u = Fx, where F was found in (a). Derive the state-space description of the closed-loop system. If w = Fx + r, what is the transfer function between y and r? (d) For x(0) = [-0.6,0.4]^ and x(0) = [0,0]^ plot x\t),x{t),y{t) and u{t) of the closed-loop system obtained in (c) and comment on your results. Use a = 1,2,5, and 10 and comment on the effects on the system response. Remark: This exercise illustrates the deterioration of system response when state observers are used to generate the state estimate that is used in the feedback control law.
4.15. Consider the system x =
4.16. Consider the system X ( ^ + 1 ) : --Ax{k)+Bu{k)+Eq{k),
y{k)=Cx{k),
where q{k) G R^ is some disturbance vector. It is desirable to completely eliminate the effects of q(k) on the output y{k). This can happen only when E satisfies certain conditions. Presently, it is assumed that q{k) is an arbitrary r x 1 vector. (a) Express the required conditions on E in terms of the observability matrix of the system. (b) If A =
,C = [1,1], characterize all E that satisfy these conditions.
(c) Suppose E eR^^^,C eRP^^, and q{k) is a step, and let the objective be to asymptotically reduce the effects of q on the output. Note that this specification is not as strict as in (a), and in general it is more easily satisfied. Use z-transforms to derive conditions for this to happen. Hint: Express the conditions in terms of poles and zeros of {A, E,C}.
4.17. Consider the system {A, B, C, D} given hy x = Ax + Bu, y = Cx + Du, where A G i?'^><«, B e R^^'P, C G 7?^><^ D G RP^'P, and detD # 0. (a) ShowthatjA,^, C, 5 } = {A - BD~^C,BD-\ of{A,B,C,D},i.e.,
-D'^Q
C(5/ - A)~^5 + D =
D~^} is the inverse system
H(s)-\
where //(i-) = CC^*/ - A) ~ ^ 5 + D is the transfer function matrix of {A, B, C, D). Verify this result using a state-space representation of the system H{s) = (s^ + l)/(^2 + 2). (b) It is possible to show (a) using results involving state feedback. In particular, let (A, B) be controllable, let u = Fx+Gr, and choose the linear state feedback control gain matrices (F,G) so that//F,GW = (C+ DF)(sI-(A +BF)y^BG + DG = L Note that this choice for F makes all eigenvalues unobservable. Use this result to prove (a). Hint: It can be shown using matrix identities that HF,G(S) = H(s) X [I-F(sI~A)-^C]-^G = H(s)[I+ F(sI-(A+ BF))-^B]G [see (2A3) to (2A5)]. Suppose now that det D = 0, but there exists a diagonal polynomial matrix X(s) = diag(pi(s)),
i = I,..., p,
with pi{s) monic stable polynomials of degree fi such that \imX{s)H{s)
= D,
where det D 9^ 0. (c) Show that a realization of X(s)H(s) is x = Ax + Bu, y = Cx + Du, i.e., the A, B are the same as above and C, D are some new matrices. (d) Determine a control law u = Fx + Gr that when applied to the system {A, B, C, D} yields HP,G{S) =
X-\sl
a diagonal transfer function matrix with stable poles at the zeros of Pi(s). Note that this is the problem of diagonal decoupling via state feedback with stability (see below and refer to Subsection 7.4D and the Notes of Chapter 7; see also Exercise 4.20). In view of the hint in (b), determine a matrix HR{S) SO that H(S)HR(S) = X-\s). The diagonal decoupling problem via state feedback is to determine a control law u = Fx -\- Gr that when applied to the system {A, B, C, D] yields a closedloop transfer function HF,G(S) that is diagonal and nonsingular. The conditions for existence of solutions can be expressed as follows: let X(s) be the diagonal matrix X(s) = diag [5^], where the nonnegative integers {f} are so that all rows of lim^^oo X(s)H(s) are constant and nonzero and let \imX(s)H(s)
= B\
Then the system can be diagonally decoupled via state feedback if and only if rank B* = p. The integers {f} are called the decoupling indices of the system. (Note that 5* = D.) It can be shown that the p X p matrix 5* can also be constructed from {A, B, C, D} as follows: if the /th row of D is nonzero, this becomes the /th row of B*. Otherwise, if f is the lowest integer, for which the /th row of CA^i'^B is nonzero, then this row becomes the /th row of B*. (e) Consider the system {A, B, C} of Exercise 4.20 and determine whether it can be diagonally decoupled via state feedback. Use both the state space matrices A, B, C and the transfer function H(s) and verify that they result to the same matrix B*.
377 CHAPTER 4 :
State Feedback and State Observers
378 Linear Systems
4.18. Consider the system {A, B, C, D} given by x = Ax + Bu, y = Cx + Du with detD ^ 0, where (A, B) is controllable and (A, C) is observable. If {A, B, C, D} = {A - BD~^C, BD~^, -D'^C, D'^} is its inverse system, show that the zeros of {A, B, C, D} are thepoles of {A, 5, C, 5 } . Also, show that the poles of {A, B, C, D} are the zeros of {A, 5, C, 5 } . 4.19. Consider the system x = Ax + Bu, y = Cx + Du, where A =
0 0 0 0
B=
1 0 0 1
c=
1 0 1 2
D=
1 0 0 1
Let u = Fx + rbea. linear state feedback control law. (a) Determine F so that the eigenvalues of A + BF are - 1 , - 2 and are unobservable from y. What is the closed loop transfer function Hf(s)(y = Hpf) in this case? Hint: Select the eigenvalues and eigenvectors of A + BF. (b) Using the Structure Theorem of Chapter 3, verify that N{s)D~^{s) -s+ 1 0 5+ 1 0 5 0 = H(s) = C(sl ~ A)-^B + D. 1 5+ 2 0 s 5+ 2 Express your results in (a) as//(5)M(5) = iN(s)D-\s))(D(s)D^\s)) Hf(s) (see Subsection 4.2D).
=
N(s)Dp\s)
4.20. Consider the system x = Ax + Bu, y = Cx, where
r 0 1 0] 0 0 0 . - 1 0 1.
ro 0] B= 1 0
c=
Lo 1.
fl 1
1
.1 0
0 -1
Note that 5+ 1
H{s) = N{s:)D-\s) = \
1
Is it possible to determine the pair (F, G)mu
HFAS)
01 - 1 [1
0 5 1
-1
= Fx + Gr so that
5+1 (5 + 2)(5 + 3) 0
5 + IJ
If your answer is yes, determine such a pair (F, G). Note that if it is required that Hf^cis) be diagonal with poles at any stable locations, then this is the problem of diagonal decoupling via state feedback. Hint: Write H(s) = \ f.
H(s) and work with H(s) to determine (F, G) so
that Hf^cis) is diagonal with poles at desired locations. 4.21. Consider the controllable and observable SISO system x = Ax + Bu,y = Cx with H(s) = C(sl - A)~^B. (a) If A is not an eigenvalue of A, show that there exists an initial state XQ such that the response to u(t) = e^\ r > 0, is y{t) = H{X)e^\ t > 0. What happens if A is a zero of//(5)? (b) Assume that A has distinct eigenvalues. Let A be an eigenvalue of A and show that there exists an initial state XQ such that with "no input" {u{t) = 0), y{t) = ke^\ t > 0, for some k B R.
4.22.
For the discrete-time case, derive expressions that correspond to formulas (2.40) to (2.45) of Subsection 4.2D.
4.23.
Consider the controllable and observable SISO system x = Ax-\-bu-\- bw,y = ex, where w is a constant unknown disturbance modelled by w = 0. If an estimate of w,w is available, then we may attempt to cancel out the disturbance by selecting u = —w. For this, consider w to be an additional state for the original system and determine an observer for this augmented system. (a) Show that the augmented system is observable if and only if 5 = 0 is not a zero of the system [or of H{s) =c{sl—A)~^b]. Hint: Use the eigenvalue criterion for observability. (b) Assume that {A,b,c} is stable (or has been stabilized) and select the gain of the observer to be K^ = [0,^], where k e R. Show that the compensator u = —w, which asymptotically cancels out the constant disturbance, is an integral feedback compensator. (c) Consider the original system H{s) = c{sl —A)~^b and use the results of Subsection 4.4B to design an output compensator that compensates for the above constant disturbance. For simplicity assume that H{s) is stable. Hint: There must be a pole ais = 0. Note that the conditions in (a) must be satisfied.
4.24.
Show that {A + BHC,B,C} is controllable and observable for any H G RP"" if and only if {A,B,C} is controllable and observable.
4.25.
Consider the system
x{k+l)=
2
"1
I 0
x{k) + 0
0
0
0" 1 u{k),y{k)0
0 0
x{k).
Is it possible to determine a linear state feedback law u{k) = Fx{k) + r{k) so that the eigenvalues of A-\-BF remain at exactly the same locations while {A-\-BF,B, C} becomes observable? If the answer is affirmative, determine such F. 4.26.
Static, or constant output feedback u = Hy-\-r,H G R'^^P^ can be used to compensate a system X = Ax-\-Bu, y = Cx, where A e R^^^,C e R^^^ and assign the eigenvalues of A + BHC of the closed-loop system i = (A + BHC)x -\-Br, y = Cx. In contrast to state feedback compensation, in general the closed-loop eigenvalues cannot be arbitrarily assigned using H even when {A,B,C} is controllable and observable. It can be shown that one can arbitrarily assign "almost always" at least min {m-\-p — l,n) eigenvalues using //; note that when p = m = 1, only one eigenvalue can be arbitrarily assigned using H. To illustrate, (a) Let A,B be as in Exercise 4.1, let C = [1,1], and determine H so that the eigenvalues of A + BHC are at -0.1025 ± y 0.04944. (b) Let A,B be as in Exercise 4.3, let C = [1,0,1], and determine H so that the eigenvalues of A-\-BHC are all at zero. (c) For an example (of a "nongenetic" case) where fewer than p-\-m—l eigenvalues can be arbitrarily assigned, consider "0 1 0 0" "0 0" 1 0 0 0 0 0 0 A = c 0 0 1 0 1 1 0 0 0 0 0 0 0 1 and show that only 2 < p-\-m aeR.
1 = 3 eigenvalues can be assigned to i ^ / a ^
379 CHAPTER 4 :
State Feedback and State Observers
380 Linear Systems
Remark: When only some of the n eigenvalues are assigned, as was the case when using constant output feedback, care should be taken to guarantee that the remaining eigenvalues will also be stable, since H will typically shift all eigenvalues. 4.27. (Inverted pendulum) Consider the inverted pendulum described in Exercise 1.20 of Chapter 1. The linearized state-space model x Ax + Bu(dibouiy = 0) assuming the friction is zero is given by r 1 g , ^8 0 0] "0 Xi ML L ML r^i 0 0 \X2 1 0 0 X2 + = mg u, 1 X3 0 0 0 M M 1x4 1 oj .0 0 0 where xi — 6, X2 = 9, X3 = s, X4 = s, and u = ^ii, the force applied to the cart. Let L = 1 m, m = 0.1 kg, M = 1 kg, and ^ = 10 m/sec^ to obtain Xi X2 X3 X4
0 1 0 0
11 0 -1 0
0 0 0 1
ol \xi 0 \x2 + 0 Us OJ\M
-l" 0 1 0
(a) Determine the eigenvalues and eigenvectors of A. (b) Plot the states when x(0) = [0.01, 0, 0,0]^. Repeat for zero initial conditions and u the unit step. Comment on your results. (c) Determine the linear state feedback control law u = Fx + r so thai the closed-loop system eigenvalues are at - 1 0 , - 3 , - 1 , and - 0 . 5 . (d) Use the LQR formulation to determine a stabilizing linear state feedback control law u = Fx + r. Comment on your choices for the weights. (e) Letx(O)^ = [0.5, 0, 0, 0]. Repeat (b) for the closed-loop system derived in (c) and (d). 4.28. (Armature voltage-controlled dc servomotor) Consider the armature voltagecontrolled dc servomotor of Exercise 2.71 in Chapter 2. Let y = xi. (a) Design a full-order state observer with eigenvalues at - 5 , - 1 , and - 0 . 5 to derive an estimate x(t) of the state x(t). For x(Of = [77/6, 0,0], x(Of = [0,0, 0] plot the error e(t) = x(t) - x(t) for r > 0 and comment on your results. (b) Assume that the system is driven by a zero-mean Gaussian, white-noise w and the measurement noise v is also zero-mean Gaussian, white noise, where w and v are uncorrected in time with covariances W = 10"^ and V = 10~^, respectively. The state-space description of the system is now x = Ax + Bu + Fw, y = Cx + v, where F = [1,1, 0]^. Design an optimal observer and compare with (a). 4.29. (Automobile suspension system) Consider the automobile suspension system of Exercise 2.74 in Chapter 2. Assume that the state-space description of the system is i: = Ax + Bu + Tw,y = Cx + v with A, B from Exercise 2.74, C = [0,0, 1, 0], and F = [ - 1 , 0, 0, 0]"^. Both process noise w and measurement noise v are assumed to be uncorrelated, zero-mean Gaussian, stochastic processes with covariances W = Ix 10""^ and V = 10~^, respectively. Let the damping constant c = 750 N sec/m and the velocity of the car V = 18 m/sec. (a) Design an optimal LQG observer-based dynamic controller. Comment on your choice of the weights. (b) Using the controller from (a), plot the states x(t) and the control input u(t) fort ^ 0 when x(0) = [1, 0, 0, 0]^, w = 0, v = 0, and the reference input of the system is r(t) = I sin(27rvr/20).
4.30. (Aircraft dynamics) Consider the systems describing the aircraft dynamics in Exercise 2.76 in Chapter 2. (a) For the state-space representation of the longitudinal motion of the fighter AFTI16, design a linear state feedback control law u = Fx + r so that the closed-loop system eigenvalues are at -1.25 ± 7*2.2651 and -0.01 ± j0.095. (b) Let y = Cx with C = [0,0, 1, 0]. Design a full-order state observer with eigenvalues at 0, -0.421, -0.587, and - 1 . (c) Let the system be compensated via the state feedback control law u = Fx + r, where x is the output of the state estimator. Derive the state-space representation and the transfer function between y and r of the compensated system. Is the system fully controllable from r? Explain. (d) Use the LQR formulation to determine a stabilizing linear state feedback control law u = Fx + r. Comment on your choices for the weights. (e) Assume that process noise w and measurement noise v are present and that both are uncorrelated, zero-mean Gaussian, stochastic processes with covariances W = 10""^ and V = 10~^, respectively. Let T = [0, 1, 1, 0]^ and design an optimal observer. (f) Design an optimal LQG observer-based dynamic controller and determine the eigenvalues of the closed-loop system. Discuss your answer in view of the results in (c). 4.31. (Chemical reaction process) [C. E. Rohrs, J. M. Melsa, and D. G. Schultz, Linear Control Systems, McGraw-Hill, 1993, p. 70.] Consider the process depicted schematically in Fig. 4.8. A reaction tank of volume V = 5, 000 gallons accepts a feed of reactant that contains a substance A in concentration CA,O- The feed enters at a rate of F gallons per hour and at a temperature TQ. In the tank some of the reactant A is turned into the desired product B. The output product is removed from the tank at the same rate that the feed enters the tank. The mixture in the tank has a uniform concentration of A, CA and a uniform temperature T.
Temperature
Product
FIGURE 4.8 Chemical reaction process
381 CHAPTER 4 :
State Feedback and State Observers
382 Linear Systems
The temperature of the water in the jacket is assumed uniform at Tj. The temperature of the water flowing into the jacket is TIQ. The system is controlled by measuring the temperature T in the tank and controlling the flow of the water in the jacket Fj by activating a valve. The equations describing the evolution of CA, T, and Tj are CA
CA,Q
t
To-
•
r-y^iCA^~^^2/n V
-T-hhCAC-^^^'^^-hiT-Tj)
TJ = TT'^J'^ ' yj
V
-TJ
+ k5(T -
TJ),
where CA,O = 0.5, TQ = 70°F, Tj^o = 70°F, and k\, k2, fe, k4, ks are appropriate constants. A linearized state-space model x = Ax + Bu, y = Cx around the equilibrium point CA = 0.245, f = 140, fj = 93.3 is given by 'xi
=
X2
>3.
'-1.1 696 0
- 22.13 X 10-4 2.9 6.5
0 1 2.4 U2 -19.5J L-^3.
pr + "
0 " 0 .-0.16_
'xi'
y = [0,1,0]
X2 .•^3.
where x\ = 8CA, X2 = ST, X3 = STj, and u = 8Fj. (a) Determine the eigenvalues and eigenvectors of A. Is the system asymptotically stable? Explain. (b) Let the input be the unit step indicating that the cooling flow is increased and held at the new value. Plot the states for /^ > 0 assuming zero initial conditions (equilibrium values). (c) Plot the states for f > 0 for zero input but with an initial concentration of substance A slightly larger than the equilibrium value, namely, x(0) = [0.1, 0, 0]^. (d) Determine the linear state feedback control law u = Fx + r so that the closed-loop system eigenvalues are at - 5 , - 1 0 , and - 1 0 . (e) Use the LQR formulation to derive a stabilizing linear state feedback control law u = Fx + r. Comment on your choices of the weights. (f) Repeat (b) and (c) for the closed-loop system derived in (c) and (d). 4.32. (Economic model for national income) Consider the economic model for national income in Exercise 2.68 in Chapter 2. (a) In which cases, (i), (ii), or (iii), is the system reachable? (b) For case (i), design a linear state feedback control law to place both eigenvalues at zero. This corresponds to a strategy for government spending that will return deviations in consumer expenditure and private investment to zero. 4.33. (Read/write head of a hard disk) Consider the discrete-time model of the read/write head of a hard disk described in Exercise 2.77 of Chapter 2. (a) Find a linear state feedback control law to assign both eigenvalues at zero. Plot the response of the discrete-time closed-loop system to a unit step and comment on your results. (b) Let y(k) = 6(k) be the position of the head at time k. Design an appropriate observer of the state and use it together with the control law determined in (a). Plot the response to a unit step and compare your results to the results in (a).
CHAPTERS
Realization Theory and Algorithms
When a linear system is described by an internal description it is straightforward to derive its external description. In particular, given a state-space description for a linear system, the impulse response, and also the transfer function in the case of time-invariant systems, were readily expressed in terms of the state-space coefficient matrices in previous chapters. In this chapter the inverse problem is being addressed: given an external description of a linear system, specifically, its transfer function or its impulse response, determine an internal, state-space description for the system that generates the given transfer function. This is the problem of system realization. The name reflects the fact that if a (continuous-time) state-space description is known, an operational amplifier circuit can be built in a straightforward manner to realize (actually simulate) the system response. The ability of reaUzing systems that exhibit desired input-output behavior is very important in applications. In the design of a system (such as a controller or a filter), the desired system behavior is frequently specified in terms of its transfer function, which is typically obtained by some desired frequency response. One must then implement a system by hardware or software that exhibits the desired input-output behavior described by the transfer function. This in effect corresponds to building a system by combining typically less complex systems in parallel, feedback, or cascade configurations. In terms of block diagrams, this corresponds to building a system by combining simpler blocks. There are of course many ways, an infinite number in fact, of realizing a given transfer function. Presently, we are interested in realizations that contain the least possible number of energy or memory storage elements, i.e., in realizations of least order (in terms of differential or difference equations). To accomplish this, the concepts of controllability and observability play a central role. Indeed, it turns out that realizations of transfer functions of least order are both controllable and observable. The theory of realizations presented in this chapter also sheds light on the behavior of systems that are built by interconnecting several other systems, as for example, in feedback control systems. In such systems, possible 383
384 Linear Systems
pole-zero cancellations between transfer functions of different subsystems can be studied, using internal descriptions and the notion of controllability and observability (see also Subsections 7.3B and 7.3C in Chapter 7).
5.1 INTRODUCTION The goal of this chapter is to introduce the theory of realization, to establish fundamental existence and minimality results, and to develop several realization algorithms. The emphasis is on time-invariant, continuous-time, and discrete-time systems. In what follows, we first provide a glimpse of the contents of this chapter. This is followed by some guidelines for the reader.
A. Chapter Description In Section 5.2, the problem of system realization is introduced, both for continuoustime and discrete-time systems. State-space realizations of impulse and pulse responses for time-varying and time-invariant systems and of transfer functions (in the case of time-invariant systems) are discussed. In Section 5.3, the existence of state-space realizations is considered first, and results are presented for both time-varying and time-invariant cases. The remainder of our development of realization theory and algorithms in this chapter concentrates primarily on time-invariant, continuous-time, and discrete-time systems. Minimal or irreducible realizations are discussed. For the time-invariant case it is shown that a state-space realization is irreducible if and only if it is both controllable and observable. Also, it is shown that if two realizations are minimal, then they must be equivalent. The order of minimal realizations is considered next, and it is shown that it can be determined directly from a given transfer function matrix without first finding a realization. This is accomplished by use of the pole polynomial of the transfer function that determines its McMillan degree and also by use of the Hankel matrix of the Markov parameters of a system. In addition, it is shown that in any minimal realization, the pole polynomial of the transfer function is the characteristic polynomial of the matrix A. In Section 5.4, a number of realization algorithms are presented. The use of duality in obtaining realizations is also highlighted. Algorithms for obtaining realizations in controller and observer form are introduced. The SISO case is treated first in the interest of clarity. Realizations with the matrix A in diagonal or block companion form are also derived. Finally, singular-value decomposition is used to obtain transfer function realizations, such as balanced realizations, in a computationally efficient manner.
B. Guidelines for the Reader In this chapter state-space realizations of input-output descriptions, which are primarily in transfer function matrix form, are developed. In the present treatment,
fundamental results are emphasized. We point out that realization theory is among the first important principal topics studied in system theory. Existence and minimality results of state-space realizations are established and several realization algorithms are developed. We note that detailed summaries of the contents of the sections are given at the beginning of Sections 5.3 and 5.4. Existence of time-invariant state-space realizations of a transfer function matrix H(s) is addressed in Subsection 5.3A, Theorem 3.3, while minimality is fully explored in two results, Theorems 3.9 and 3.10 in Subsection 5.3B. It is useful to be able to determine the order of a minimal realization directly from H(s) and this is discussed in Subsection 5.3C; we note that the pole polynomial of H{s) introduced in Section 3.5 of Chapter 3 is required in one of the approaches presented. In Section 5.4, several realization algorithms are developed. The use of duality in realizations is emphasized in Subsection 5.4A. At a first reading, the reader may concentrate on minimality of realizations in Subsection 5.3B and on the order of minimal realizations in Subsection 5.3C; then study one or two realization algorithms, e.g., the one that leads to a realization with A diagonal in Subsection 5.4C, and to a realization with A, B in controller form in Subsection 5.4B. It is also important to study Subsection 5.4A on realizations using duality. If realization algorithms with good numerical properties are of primary interest, then the reader should concentrate on Subsection 5.4E, where realizations using singular-value decomposition are presented.
5.2 STATE-SPACE REALIZATIONS OF EXTERNAL DESCRIPTIONS In this section, state-space realizations of impulse responses for time-varying and time-invariant systems and of transfer functions for time-invariant systems are introduced. Continuous-time systems are discussed first in Subsection 5.2A, followed by discrete-time systems in Subsection 5.2B.
A. Continuous-Time Systems Before formally defining the problem of system realization, we first review some of the relations that were derived in Chapter 2. We consider a system described by equations of the form X = A(t)x + B(t)u,
y = C{t)x + D{t)u,
(2.1)
where A{t) E T^^^^^ B{t) G R'''''^, C{t) G i^^^", and D{t) G /^^^^ are continuous matrices over some open time interval {a, b). The response of this system is given by y{t) = C{t)(S^{t,tQ)xo +
H{t,T)u{T)dT,
(2.2)
where ^{t, to) is the nX n state transition matrix of i: = A(t)x, x(to) = XQ (the initial condition), and H(t, r) is the p X m impulse response matrix of this system, given
385 CHAPTER 5 :
Realization Theory and Algorithms
386 Linear Systems
by H{t,T) =
C(t)^(t,
T)B(T)
+ D(t)8(t -
for t > r,
T),
0,
(2.3)
for t < T.
In the time-invariant case, (2.1) assumes the form X = Ax + Bu,
y = Cx + Du,
(2.4)
and the system response is in this case given by y(t) = Ce^^xo +
H(t, T)u(r) dr,
(2.5)
where, without loss of generality, ^o was taken to be zero. The impulse response is now given by the expression H{t,
Ce^^'-^^B + DS(t - T),
for t ^ r,
0,
for t < T.
T)
(2.6)
Recall that the time invariance of system (2.4) implies that H(t, r) = H(t - r, 0), and therefore, r, which is the time at which a unit impulse input is applied to the system, can be taken to equal zero (r = 0), without loss of generality, to yield H(t, 0). The transfer function matrix of the system is the (one-sided) Laplace transform of H(t, 0), namely. H(s) = iE[H(t, 0)] = C(sl - A)~^B -h D.
(2.7)
Let {A(0, B(t), C{t\ D(t)} denote the system description (2.1) and let H(t, r) be a /? X m matrix with real functions of arguments t and r as entries. DEFINITION 2.1. A realization ofH(t, r) is any set{A(0, B(t\ C{t), D{t)}, the impulse response of which is H{t, r). That is, {A{t\ B(t\ C(t\ D(t)} is a reahzation of H(t, r) if (2.3) is satisfied. (See Fig. 5.1.) •
D{t)
U(fo) u{t)
B{t)
+ ^7^ '^^\ \ J n
+
x{t)
C{t)
ys>^
+ \
A{t)
FIGURE 5.1 Block diagram reahzation of {A(t\ B{t), C{t), D(t)} Note that it is not necessary that any given pX m matrix H(t, r) be the impulse response to some system of the form (2.1). The conditions on H(t, r) under which a realization {A(t), B{t), C{t), D(t)} exists are given in the next section.
In the time-invariant case, a realization is commonly defined in terms of the transfer function matrix. We let {A, B, C, D] denote the system description given in (2.4) and we let H{s) be a p X m matrix with entries that are functions of s.
387 CHAPTERS:
Realization Theory and DEFINITION 2.2. A realization of H{s) is any set {A, B, C, D}, the transfer function Algorithms matrix of which is H{s), i.e., {A, B, C, D] is a realization of H{s) if (2.7) is satisfied. •
As will be shown in the next section, given H{s), a condition for a realization {A, B, C, D} of H{s) to exist is that all entries in H{s) are proper, rational functions. Alternative conditions under which a given set {A, B, C, D] is a realization of some H{s) can easily be derived. To this end, we expand ^(5") in a Laurent series to obtain (2.8)
H{s) - //o + ^ 1 ^ " + ^2^~ + DEFINITION 2.3. The terms Hu i = 0,1, 2, the system.
in (2.8) are the Markov parameters of
The Markov parameters can be determined by the formulas HQ = lim H(sl
Hi = lim s(H(s) - HQI
H2 = lim s^(H(s) -Ho-
His'^l
and so forth. Recall that relations involving the Markov parameters were alluded to earlier in Exercise 2.63 of Chapter 2. THEOREM 2.1. The set {A, B, C, D) is a realization of ^(^) if and only if if0 - ^
and
Hi = CA^'^B,
i=
12,....
(2.9)
Proof, H(s) = D + C(sl - Ay^B = D + Cs'^I - s'^Ay^B = D+ Cs-'{Y.U{s-'Am = D + Y.U[CA^-'B]s- , from which (2.9) is derived in view of (2.8). By definition, the impulse response description of a linear system contains no information about the initial conditions, or the initial energy stored in the system. In fact, H{t, T) is determined by assuming that the system is at rest before r, the time when the impulse input 8{t - T) is applied. It is therefore apparent that different state-space realizations of H{t, r) will yield the same zero-state response, while their zero-input response, which depends on initial conditions, can be quite different. Note that if a realization of a given H{t, r) exists, then there are infinitely many realizations. Given {A{t), B(t), C{t), D(t)}, a realization of H(t, r), other realizations with the same dimension n of the state vector can readily be generated by means of equivalence transformations. Recall from Chapter 2 that equivalent representations generate the same impulse response. To illustrate, in Example 3.1 of Section 5.3, the system i: - P(t)N(t)u(t) mdy(t)
= M(t)P~\t)x(t)
with N(t) = '
-t
and M(t)
[t, 1] is a realization of H(t, r) = t - r for any P(t) such that P ^(t) and P(t) exist and are continuous.
388 Linear Systems
B. Discrete-Time Systems The problem of realization in the discrete-time case is defined as in the continuoustime case: given an external description, either the unit pulse [(discrete) impulse] response H{k, €), or in the time-invariant case, typically the transfer function matrix H{z) (see Chapter 2), determine an internal state-space description, the pulse response of which is the given H{k, £). The realization theory in the discrete-time case essentially parallels the continuous-time case. There are of course certain notable differences because in the present case the realizations are difference equations instead of differential equations. We point to these differences in the subsequent sections. Some of the relations derived in Section 2.7 of Chapter 2 will be recalled next. We consider systems described by equations of the form x(k + 1) - A(k)x(k) + B(k)u(k\
y(k) = C(k)x(k) + D(k)u(k),
where A(k) G 7?"><", B(k) G J^^X'^, C(k) G /?^x^ and D(k) G of this system is given by the expression
RP'''^,
(2.10)
The response
k-i
y(k) = C(k)(k ko)xQ + ^
H(k i)u{i),
k > k^,
(2.11)
where ^{k, ko) denotes the n X n state transition matrix of the system x(k + 1) = A(k)x(kX x(ko) = xo is the initial condition, and H(k, i) is the p X m pulse response matrix given by C(k)<^(k, i + l)B{i\
k > /,
H(k i) = < D(k),
k = /,
0,
(2.12)
k< i.
The state transition matrix can readily be determined to be ^{k,^
{ A{k= \
l)A(fc-2)---A(€),
yt>€,
[ /,
(2.13)
k= t
In the time-invariant case, (2.10) assumes the form x{k + 1) - Ax{k) + Bu{k),
y(k) = Cx(k) + Du(kl
(2.14)
and the system response of (2.14) is given by k-l
y(k) = CA^xo + X ^ ( ^ ' ^*)"(^*)^
^ ^ 0,
(2.15)
where, without loss of generality, ^o was taken to be zero. The pulse response is now given by
CA'' -(i+i)B,
k> i,
Hik, i) = < D,
k = /,
0,
k< i.
(2.16)
Recall that since the system (2.14) is time-invariant, H{k, i) = H(k - i, 0) and /, the time the pulse input is applied, can be taken to be zero, to yield H(k,0) as the external system description. The transfer function matrix for (2.14) is now the (onesided) z-transform of H(k, 0). We have. H(z) = %{H{k, 0)} = C(zl - Ay^B
+ D.
(2.17)
Now let {A{k), B(k), C(k), D(k)} denote the system description (2.10) and let H(k, /) be a /7 X m matrix with real function entries of arguments k and / defined for k^ /. DEFINITION 2.4. A realization of H(k, i) is any set of matrices {A(k), B(k), C(k), D(k)}, the pulse response of which is H(k, i). • Let H(z) be a /? X m matrix with functions of z as entries. DEFINITION 2.5. A realization of H{z) is any set {A, B, C, D], the transfer function matrix of which is H{z). • A result that is analogous to Theorem 2.1 is also valid in the discrete-time case [with H{s) replaced by H{z)\ (show this). The remarks following Theorem 2.1 concerning the zero-state response of a system and the uniqueness of realizations are also valid in the present case. Thus, all realizations of the pulse response or the transfer function will yield the same zero-state response, while their zero-input response, which depends on initial conditions, can be quite different. Also, if a realization of H{k, i) or H{z) exists, then there exists an infinite number of realizations (show this).
5.3 EXISTENCE AND MINIMALITY OF REALIZATIONS In this section the existence and minimality of internal state-space realizations of a given external description are determined. The existence of realizations is examined first. Given 3. p X m matrix H(t, r), conditions under which this matrix is the impulse response of a linear system described by equations of the form i: = A(t)x + B(t)u,y = C(0^ + ^ ( 0 ^ are given in Theorem 3.1. Theorem 3.2 provides the corresponding result for time-invariant systems. For such systems, H(s) is typically given in place of H(t, r), and conditions for H(s) to be the transfer function matrix of a system described by equations of the form X = Ax + Bu, y = Cx -\- Du are given in Theorem 3.3. It is shown that such realizations exist if and only if H(s) is a matrix of rational functions with the property that lims-^oo H(s) is finite. The corresponding results for discrete-time systems are then developed and presented in Theorems 3.5, 3.6, and 3.7. Realizations of least order, also called minimal or irreducible realizations, are of interest to us since they realize a system, using the least number of dynamical elements (minimum number of elements with memory). The main emphasis in this section is on time-invariant systems and realizations of transfer function matrices H(s). The principal results are given in Theorems 3.9 and 3.10, where it is shown that minimal realizations are controllable (-from-the-origin) and observable and that all minimal realizations of H(s) are equivalent representations. The order of any minimal realization can be determined directly without first determining a minimal realization, and this can be accomplished by using the characteristic polynomial
389 CHAPTER 5 :
Realization Theory and Algorithms
390 Linear Systems
and the degree of H(s) (Theorem 3.11) or from the rank of a Hankel matrix (Theorem 3.13). All the results on minimality of realizations apply to the discrete-time case as well with no substantial changes. This is discussed at the end of the section. A. Existence of Realizations Continuous-time systems Let the /? X m matrix H(t, r) with t,T E (a, b) be given. A realization of H(t, r) was defined in the previous section as an internal description of the form (2.1) denoted by {A{t), B{t), C{t), D(t)}, the impulse response of which is H(t, r). THEOREM 3.1. H(t, T) is realizable as the impulse response of a system described by (2.1) if and only if H(t, r) can be decomposed into the form H(t, T) = M(t)N(T) + D(t)8(t - r)
(3.1)
for r > T, where M, A'', and D SLYQ p X n, n X m, and p X m matrices, respectively, with continuous real-valued entries and with n finite. Proof. (Sufficiency) Assume the decomposition (3.1) is true and consider the realization {0, A^(0, M(t\ D(t)}, which yields the system x = N(t)u(tX y(t) = M(t)x(t) + D(t)u(t). The state transition matrix is ^(t, T) = /, since the homogeneous equation is in this case X = A(t)x = 0. Applying (2.3), we obtain as the impulse response, H(t, r) = M(t) • / • N(T) + D(t)8(t - r) for t > r, and H(t, T) = Ofovt < r, which equals H(t, r). (Necessity) Let (2.1) be a realization of H(t, r). Then (2.3) is true, and in view of the identity (/, r) = <|)(/, (T)^((7, T), H(t, r) can be written for r > r as H(t, r) = [C(t)<^(t, a)][(a, T)B(T)] + D(t)8(t - T) = M(t)N(T) + D(t)8(t - r), where M(t) = C(t)(!?(t, a) and N(r) = (t>(a, T)B(T) (a fixed). Therefore, the decomposition (3.1) is necessary. • Since the matrices in (2.1) are taken to be continuous, we require in Theorem 3.1 that M(t), N(t), and D(t) be continuous, although this restriction is not necessary. EXAMPLE 3.1. Let H(t, T) = t - T, t > T. Then H(t, r) = [t, 1] Therefore, x
1 -t
n=
M(t)N(T).
u(t), y(t) = [t, l]x(t) is a reahzation.
EXAMPLE 3.2. Let H(t, r) = l/(t - T). In this case a decomposition of the form (3.1) does not exist, and therefore, H(t, r) does not have a realization of the form (2.1). • If a system is time-invariant and is described by (2.4), its impulse response is given by (2.6). The following result establishes necessary and sufficient conditions for the existence of time-invariant realizations. THEOREM 3.2. H(t, T) is realizable as the impulse response of a system described by (2.4) if and only if H(t, r) can be decomposed for ^ > r into the form H(t, T) = M(t)N(T) + D(t)8(t - T),
(3.2)
where M(t) and N(t) are differentiable and H(t, T) = H(t - T, 0).
(3.3)
The first part of this result is identical to Theorem 3.1. For time invariance, we require in addition relation (3.3) and differentiability.
Proof, {Necessity) Let (2.4) be a realization. Then in view of (2.6), H{t, r) = H(t T,0) = Ce^^'-'^B + D8{t - r) = {Ce^'Xe'^-'B) + Dd{t - r). Let M(t) = Ce^\ N(T) = e'^^B, and note that both M(t) and N(t) are differentiable. (Sufficiency) The proof of this part is much more involved. The complete proof can be found, for example, in Brockett [1], p. 99. In the following, we give an outline of the proof (a proof by construction). First, it can be shown that given H(t, r) with decomposion (3.1), a realization of the form x = N(t)u(t\ y(t) = M(t)x(t) + D(t)u(t) can be found where n, the dimension of the state vector, is the smallest possible. Note that this system is controllable and observable. Now define W(to, ti) = j / ^ N(a)N^(a)da and Wi(to, h) = j^^^ [(d/da)N(a)] N^(a)da. Then, using H(t, r) = H(t - T, 0), it can be shown that {A, B, C, D) - {-Wi(tQ, h)W~\tQ, h), N(0), M(0), D} is a realization of H(t, T). Note that W~H^> ^i) exists because {0, N{t), M{t), D] was taken to be of minimal dimension. Therefore, this system is controllable. • EXAMPLE 3.3. Consider again H{t, r) = t - T given in Example 3.1, where a timevarying realization was derived. This H(t, r) certainly satisfies the conditions of Theorem 3.2, and thus, a time-invariant realization x = Ax + Bu, y = Cx + Du also exists. Unfortunately, the proof of Theorem 3.2 does not provide us with a means of deriving such realizations. In general it is easier to consider the Laplace transform of H(t, r) and use the algorithms that we will develop in Section 5.4 to show that a time-invariant realization of H(t, T) = t - T is given by i: =
"0 1 .0 0.
X+
"0" u, y = [1, 0]x. (Verify this 1.
and also see Example 3.4.) In the remainder of this chapter we shall concentrate primarily on time-invariant realizations of impulse responses. In fact we shall assume that the system transfer function matrix H(s), which is the Laplace transform of the impulse response, is given and we shall introduce methods for deriving realizations directly from H(s). Given a pX m matrix H(s), the following result establishes necessary and sufficient conditions for the existence of time-invariant realizations. THEOREM 3.3. H(s) is reahzable as the transfer function matrix of a time-invariant system described by (2.4) if and only if H(s) is a matrix of rational functions and satisfies \im H(s) < CO,
(3.4)
i.e., if and only if H(s) is a. proper rational matrix. Proof (Necessity) If the system x = Ax -\- Bu, y = Cx + Du is Si reahzation of H(s), then C(sl - A)~^B -\- D = H(s), which shows that H(s) must be a rational matrix. Furthermore, \im H(s) = D,
(3.5)
which is a real finite matrix. (Sufficiency) lfH(s) is a proper rational matrix, then any of the algorithms discussed in the next section can be applied to derive a realization. • EXAMPLE 3.4. Loi H(s) = l/s^,(y(s) = //(^)w(^)), which is the transfer function of the double integrator. Then, using the controller form realization algorithm of Section 5.4, a realization of H(s) is given by i: = ^-^[H(s)]
= ^-^[l/s^]
"0 1" 0 X+ u,y = [1,0]x. Notice that .0 0. 1.
= t = H(t, 0), the same as in Examples 3.1 and 3.3.
391 CHAPTER 5 :
Realization Theory and Algorithms
392 Linear Systems
Theorem 3.3 can be used to show what types of entries are required for H{t, 0) for it to be reaUzable as a Hnear time-invariant continuous-time system of the form given in (2.4). In particular, H{t, 0) = X~^[H{sy\ and the fact that all entries of H{s) are proper rational functions (Theorem 3.3) implies the next result. COROLLARY 3.4. A pX m matrix H{i){H{t, 0)) is realizable as the impulse response of a system described by equations of the form (2.4) if and only if all entries of H{t) are sums of terms of the form at^e^^ and I38(t), where a, /3 are real numbers, k is an integer (k > 0), and A is a complex scalar. Proof. The proof is left as an exercise for the reader. (Refer to Chapter 2, Subsection 2.4B for a review of Laplace transforms. Also, refer to the discussion of modes and asymptotic behavior of a system in that chapter.) • EXAMPLE 3.5. Let H(t) = W + e~^ + ^(0. e% In view of the above corollary, H{t) is realizable as the impulse response of a system described by (2.4). In fact, H{s) = ^{H(t)} =
's^ + 2s-l ^2-1
1 " and the system i: = Ax + Bu,y = Cx + DwwithA = ' s-l
"2 1 r - 1 2" ,B = ,C = [1,0], D = [1, 0] is a reahzation (verify this). Comparing [l .1 1. L 0 i_ with Theorem 3.2, we write H(t -T,0)
= [e'- + e-^'-^^ + 8(t - T), e'-'] [e-\ e' - e-n e^ + e~ M(t)N(T) + D(t)8(t
+ [l,0]5(r-T) T),
which shows that the given H{t) [resp. H{t, 0)] is indeed realizable by a system described by (2.4). Actually, in this case M{t) and N{T) were chosen so that M{t) = Ce^^ and N(T) = e ^^B, where e^^
0
e'
(verify this).
Discrete-time systems Results for the existence of realizations of discrete-time systems that are analogous to the continuous-time case can also be established. In the following result, which corresponds to Theorem 3.1, H(k, i), k > /, denotes a. p X m matrix. THEOREM 3.5. H(k,i)is realizable as the pulse response of a system described by (2.10) if and only if H(k, i) can be decomposed into the form M(k)N(i),
k> /,
D(k),
k = i.
H(k, i)
(3.6)
Proof, {Sufficiency) We consider the reahzation {/, N{k), M(k), D(k)}, i.e., x(k + 1) = x(k) + N(k)u(k) and y(k) = M(k)x(k) + D(k)u(k). In this case, ^(k, ^ = I,k> t Applying (2.12), it is immediately verified that the pulse response of this system is the given H(k, i). {Necessity) For any realization (2.10), C(k)<^(k,i + l)B(i) = C(k)A(k - 1 ) . . . A(i + l)B(i) = C(k)A(k - l)...A(a)A(a - l)...A(i + l)B(i) = (C(k)^(k,a)) X (0(o-, / + l)B(i)) = M(k)N(i), i < a ^ k, which in view of (2.10), implies (3.6). • The discrete-time system result corresponding to Theorem 3.2 for time-invariant systems is considered next.
THEOREM 3.6. H(kJ) is realizable as the pulse response of a system described by (2.14) if and only if H(k, i) can be decomposed as in (3.6), and (3.7)
H(k, i) = H(k - i, 0).
Proof, (Necessity) This part of the proof is the same as the necessity proof of the previous theorem. Also, in view of (2.16), it is clear that H(k, i) = H(k - i, 0). (Sufficiency) Let H(k, i) satisfy (3.6) and (3.7). The proof is by construction, considering a least-order realization x(k + 1) = x(k) + N(k)u(k), y(k) = M(k)x(k) + D(k)u(k) and proceeding along similar lines as in the proof of Theorem 3.2. • EXAMPLE 3.6. Let H(k, i) = k-
/, k > /. Here H(k, i) = [k, 1]
=
M(k)N(i),
k > /, and H(k, i) = 0 = D(k), k = i. In view of Theorem 3.5, there exists a realization, the pulse response of which is the given H(k, i). A particular reahzation is given 1 by the system equations x(k + 1) = x(k) + u(k), y(k) = [k, l]x(k) (see the proof -k of Theorem 3.5). (Verify this.) This is of course a time-varying reahzation. However, here H(k, i) = H(k - i, 0), which in view of Theorem 3.6, implies that a time-invariant realization also exists. Such a realization is x(k + 1) = Ax(k) + Bu(k), y(k) = Cx(k), where A
B =
C = [0, 1]. [Verify that in the present case H(k, 0) =
CA^ ^B,k> 0.] This realization was determined using H(z) = ^H(k, 0) and the controller form realization algorithm in Subsection 5.4B.
zl(z - If •
Given apXm matrix H(z), the next theorem establishes necessary and sufficient conditions for time-invariant realizations. This result corresponds to Theorem 3.3 for the continuous-time case. Notice that the conditions in these results are identical. THEOREM 3.7. H(z) is realizable as the transfer function matrix of a time-invariant system described by (2.14) if and only if H(z) is a matrix of rational functions and satisfies the condition that (3.8)
lim H(z) < 00.
Proof
Similar to the proof of Theorem 3.3.
COROLLARY 3.8. A pX m matrix H(k) [resp., H(k, 0)], A: > 0, is realizable as the pulse response of a system described by (2.14) if and only if all entries of H(k) are sums of polynomial terms of the form a\^, I3k(k — 1)- • -(A: - € + 1)A^~^, and y, where a, p, y denote real numbers, k, £ are nonnegative integers, and A is a complex scalar. Proof The details of the proof are left to the reader. Note that H(k, 0) = ^-^[H(z)\ where in view of Theorem 3.7, all the entries of H(z) are proper rational functions. Refer to Section 2.7 of Chapter 2 for a review of z-transforms and a discussion of the modes and the asymptotic behavior of discrete-time systems. • EXAMPLE 3.7. Let H(z) = zl(z - 1)^. In view of Theorem 3.7, H(z) is realizable as the transfer function of a system described by (2.14). Such a realization is given by x(k +1) = Ax(k) + Bu(k), y(k) = Cx(k) + Du(k), with A =
-1
2r-[i:Ic =
[0,1], D = 0 (refer to Example 3.6). Note that in this case, H(k, 0) = ^-^[H(z)] = k, k> 0, which, in view of Corollary 3.8, imphes that H(k, 0) = kis realizable as the pulse response of a system described by (2.14). Such a reahzation is given above. •
393 CHAPTER 5 :
Realization Theory and Algorithms
394 Linear Systems
B. Minimality of Realizations As was discussed in Section 5.2, realizations of an impulse response H{t, r) can be expected to generate only the zero-state response of a system, since the external description H{t, r) has, by definition, no information about the initial conditions and the zero-input response of the system. A second important point to take note of is the fact that if a realization of a given H(t, T) exists, then there exist an infinite number of realizations. It was pointed out that if (2.1), denoted by {A{t\ B(t), C{t\ D{t)}, is a reaUzation of the /? X m matrix H{t, T), then realizations of the same order n, i.e., of the same dimension n of the state vector, can readily be generated by an equivalence transformation. Recall that in Subsection 2.6C of Chapter 2 it was shown that equivalent state-space (internal) representations generate identical impulse responses (external representations). There are, of course, other ways of generating alternative realizations. In particular, if (2.1) is a reahzation of H(t, r), then, for example, the system X = A{t)x + B(t)u,
y = C(t)x + D(t)u
z = F(t)z + G(t)u is also a realization. This was accomplished by adding to (2.1) a state equation z = F(t)z + G(t)u that does not affect the system output. The dimension ofF, dim F, and consequently the order of the realization, n + dim F, can be larger than any given finite number. In other words, there may be no upper bound to the order of the realizations of a given H(t, r). There exists, however, a lower bound, and a realization of such lowest order is called a least-order minimal or irreducible realization. DEFINITION 3.1. A realization X = A(t)x + B(t)u,
y = C(t)x + D(t)u
(3.10)
of the impulse response H(t, r) of least order n (A(t) E /?"><") is called a least order, or a minimal order, or an irreducible realization of H(t, r). • If the given impulse response H(t, r) satisfies the conditions of Theorem 3.2, so that a time-invariant realization exists, then one usually talks about minimal or irreducible realizations of tho p X m transfer function matrix H{s) = i£[H(t, 0)], and this is the case on which we shall concentrate in the remainder of this chapter. It should be noted that results corresponding to Theorems 3.9 and 3.10 exist for timevarying systems as well. The interested reader is encouraged to consult, for instance, Brockett [1] for such results. Briefly, as will be shown for the time-invariant case, system (3.10) is a minimal realization of H(t, r) if and only if the representation is controllable (-from-the-origin or reachable) and observable. Note that in this chapter the term controllable is used in place of reachable, to conform with accepted use in the literature. By controllability we will really mean controllability-from-the-origin or reachability. This distinction is not important in continuous-time systems, but it is important in discrete-time systems. DEFINITION 3.2. A realization x = Ax + Bu,
y = Cx + Du
(3.11)
of the transfer function matrix H{s) of least order n{A G /?«><«) is called a least-order, or a minimal, or an irreducible realization ofH{s). •
Theorems 3.9 and 3.10 completely solve the minimal realization problem. The first of these results shows that a realization is minimal if and only if it is controllable (-from-the-origin or reachable) and observable, while the second result shows that if a minimal realization has been found, then all other minimal realizations can be obtained from the determined realization, using equivalence of representations. Controllability (-from-the-origin, or reachabihty) and observabiHty play an important role in the minimality of realizations. This is to be expected, since these properties, as was discussed at length in Chapter 3, characterize the strength of the connections between input and state, and between state and output, respectively. Therefore, it is reasonable to expect that they will play a significant role in the relation between internal and external descriptions of systems. Indeed, it was shown in Subsection 3.4C that only that part of a system that is both controllable and observable appears in H(s). In other words, H(s) contains no information about the uncontrollable and/or unobservable parts of the system. To illustrate this, consider the following specific case. EXAMPLE 3.8. L e t / t( s) = 1/ is + 1). Four different re
(i) {A =
"0 1 "0 ,B = IC .1 .1 0.
= [-HID
= 0}.
-1 "0 1 , C = [0, 1], D = 0}. ,B = 1. .1 0. "1 0 " 0 , C = [0, 1], D = 0}. (iii) {A = P - 1 . ,B = 1 (ii) {A =
(iv) {A = -IB
=
ic
1,D = 0}.
The eigenvalue +1 that in (i) is unobservable, in (ii) is uncontrollable, and in (iii) is both uncontrollable and unobservable does not appear in H(s) at all. Realization (iv), which is of order 1, is a minimal realization. It is controllable and observable. • THEOREM 3.9. An /2-dimensional realization {A, B, C, D} of H(s) is minimal (irreducible, of least order) if and only if it is both controllable and observable. Proof, (Necessity) Assume that {A, B, C, D} is a minimal realization but is not both controllable and observable. Then, using Kalman's Canonical Decomposition of Subsection 3.4A, one may find another realization of lower dimension that is both controllable and observable. This contradicts the assumption that {A, B, C, D) is a minimal realization. Therefore, it must be both controllable and observable. {Sufficiency) Assume that the realization {A, B, C, D} is controllable and observable, but there exists another realization, say, {A, B, C, D} of order n < n. Since they are both realizations of H{s), or of the impulse response H{t, 0), then Ce^'B + D8{t) = Ce^'B + D8(t)
(3.12)
for all r > 0. Clearly, D = D = \ims-^ccH(s). Using the power series expansion of the exponential and equating coefficients of the same power of t, we obtain CA^B = CA^B,
k = 0,1,2,.
(3.13)
i.e., the Markov parameters of the two representations are the same (see Theorem 2.1 in Section 5.2). Let %n = [B, AB,..., A'^-^B] E Z^^^^'"",
395 CHAPTERS:
Realization Theory and Algorithms
C CA
396 Linear Systems
r: Dpnxn
and
(3.14)
CA n-l Then the pn x mn matrix product ^n^n assumes the form •
CA^'-^B CA^'B
CB CAB
CAB CA^B
CA^'-^B
CA^'B
CA^^'-^B
CB CAB
CAB CA^B
CA^'-^B CA^'B (3.15)
CA'^-^B
CA'^B
CA^^'-^B
In view of Sylvester's rank inequality, which relates the rank of the product of two matrices to the rank of its factors, we have rank 0^ + rank ^n — ^^ rank {^n^n) ^ rnin {rank ^„, rank ^n)
(3.16)
and we obtain that rank dn = ^ci^k "^n = ^, ^^^^ {^n^'^n) = ^- This result, however, contradicts our assumptions, since n = rank {dn'^n) ^ min(ran^ dn.rank ' ^ ) < n because n is the order of {A,5, C , 5 } . Therefore, n
B = PB,
and
c = cp-
(3.17)
Furthermore, if P exists, it is given by P = ^^(^^^)-i
or
P =
(3.18)
Proof, (Sufficiency) Let the realizations be equivalent. Since {A,B,C,D} is minimal, it is controllable and observable and its equivalent representation {A,B,C,D} is also controllable and observable, and therefore, minimal. Alternatively, since equivalence preserves the dimension of A, the equivalent realization {A,5,C,5} is also minimal.
{Necessity) Suppose {A, B, C, D] is also minimal. We shall show that it is equivalent 397 to {A, B, C, D}. Since they are both realizations of H{s), they satisfy D = D and CHAPTER 5 : Realization CA^B = CA^B, k = 0,1,2.. (3.19) Theory and as was shown in the proof of Theorem 3.9. Here, both realizations are minimal, and Algorithms therefore, they are both of the same order n and are both controllable and observable. Define ^ = ^„ and 0 = ©„, as in (3.14). Then, in view of (3.15), 0^ = 0^ and premultiplying by 0^, we obtain 0^0^ = 0^0^. Using Sylvester's inequality, we obtain rank 0^6 = n, and therefore, % = [(O^O)"^©^©]^ = P%,
(3.20)
where P = (O^©)'^©^© G 7?"><\ Note that rank P = n since rank ©^© is also equal to n, as can be seen from rank ©^©^ = n and from Sylvester's inequaUty. Therefore, P qualifies as a similarity transformation. Similarly, ©^ = ©^ implies that ©^^^ = , and 0 = ©[^^^(^^^)-i] = ©P,
(3.21)
where P = %%^{%%^y^ G T?"^'^ with rank P = n. Note that P = P. To show that P is the equivalence transformation given in (3.17), we note that ©A^ = ©A^ from (3.15). Premultiplying by ©^ and postmultiplying by ^^, we obtain PA = AF, in view of (3.20) and (3.21). To show that PB = fiandC = CP, we simply use the relations P ^ = % and © = ©P, respectively. •
C. The Order of Minimal Realizations In the next subsection algorithms to derive minimal realizations of H{s) are developed. One could ask the question whether the order of a minimal realization of H(s) can be determined directly, without having to actually derive a minimal realization. The answer to this question is yes, and in the following we will show how this can be accomplished. When deriving realizations, it is of advantage to know at the outset what the order of a minimal realization ought to be, since in this way erroneous results can be avoided by cross checking the validity of the results. Determination via the characteristic or pole polynomial of H(s) The characteristic polynomial (or pole polynomial), PH(S), of a transfer function matrix H(s) was defined in Section 3.5 using the Smith-McMillan form of H(s). The polynomial PH(S) is equal to the monic least common denominator of all nonzero minors of H(s). The minimal polynomial of a transfer function matrix H(s), mnis), was defined as the monic least common denominator of all nonzero first-order minors (entries) of H(s). DEFINITION 3.3. The McMillan degree of H(s) is the degree of PH(S).
•
The number of poles in H(s), which are defined as the zeros of PH(S) in Definition 5.1 in Section 3.5, is equal to the McMillan degree of H{s). The degree of H{s) is in fact the order of any minimal realization of H(s), as the following result shows. THEOREM 3.11. Let {A, 5, C, D} be a minimal realization of H(s). Then the characteristic polynomial of H(s\ PH{S), is equal to the characteristic polynomial of A, a{s) = \sl - A|, i.e., PH{S) = a(s). Therefore, the McMillan degree of H(s) equals the order of any minimal realization.
398 Linear Systems
Proof. The proof outlined here is based in part on results that will be established in Sections 7.2 and 7.3 of Chapter 7. (The reader may wish to postpone reading the proof until Sections 7.2 and 7.3 have been covered.) First, we note that '\fH{s) = N(s)D(s)~^ with N(s), D(s) right coprime polynomial matrices, then the characteristic or pole polynomial of H(s) is given by pnis) = k det D(s), where the real k is such that k det D(s) is monic. This can be seen, for example, by reducing H(s) to its Smith-McMillan form [see (5.3) in Chapter 3]. We have. diagi-j^,..., Ui(s)H(s)U2(s) =
-^)
0
SMH(S) ^p—r,m—r
diag{ijJx,...,\lJr)
(ei,...,€r)
0 =
N,{s)D-\s\
where r = rank H(s) and Ui, U2 are unimodular matrices. Then H = (U^^Ns) X (U2Ds)~^ = A^D~\whereMZ)arerightcoprime. Notethat A:(i^/^Z)(5) = iAi^2***^r = PH(S), by definition. Now this is true for any right coprime factorization since these are related by unimodular postmultiplication (see Subsection 7.3A in Chapter 7). The controllable and observable realization {A, B, C, D} is now equivalent to any other controllable and observable realization of the form Dz = u,y = Nz with D, N right coprime (see Subsection 7.3A). This implies that 15"/ - A| = k\D{s% since such equivalence relation preserves the system eigenvalues. Note that the same result can be derived using the Structure Theorem of Chapter 3 (show this). Therefore, the pole polynomial of H{s) is given by PH{S) = \sl ~ A\. • It can also be shown that the minimal polynomial ofH{s), mnis), is equal to the minimal polynomial of A, am(s), where {A, B, C, D] is any controllable and observable realization of H(s). This is illustrated in the following example. EXAMPLE 3.9. LQI H(s) =
[1/^
2/s^
0
-\/s\
. The first-order minors, the entries of H(s),
have denominators s, s, and s and therefore, mnis) = s. The only second-order minor is - l/s^ and PH(S) = s^ with deg PH{S) = 2. Therefore, the order of a minimal realization is 2. Such a realization is given hy x = Ax + Bu and y = Cx with A = L A,B = 0 0 "1 2" ,c = "1 0" It can be verified first that this system is a realization of H(s) and
.0
-1.
0 1_
then we verify that it is controllable and observable, and therefore, minimal. Notice that the characteristic polynomial of A is a(s) = s^ = PH{S) and its minimal polynomial is a^{s) = s = rriHis). • The above example also shows that when H{s) is expressed as a polynomial matrix N{s) divided by a polynomial, i.e., H{s) = (l/mH(s))N(s), then the roots of niH are not necessarily the eigenvalues of a minimal realization of H{s). They are in general a subset of those eigenvalues, since the minimal polynomial always divides the characteristic polynomial. In the case when H(s) is a scalar, however, the roots of rriH = PH are the eigenvalues of any minimal realization of H(s), as the next example shows.
EXAMPLE 3.10. Let H(s) = n(s)/d(s) be a scalar proper rational function. Applying Definition 5.1 of Section 3.5, we obtain PH(S) = rnnis) = d(s) and the order of a minimal realization is deg PH(S) = degd(s). Thus, given H(s) = l/(s^ + 3^ + 2), we know that a minimal realization is of second order since deg (s^ + 3^ + 2) = 2. A minimal reahzation in controller form is given by A
B
C = [1,0]. No-
tice that the minimal polynomial of A is am(s) = a(s) = s^ + 3s + 2, the characteristic polynomial of A, which is equal to d(s) = PH(S) = rnnis), as expected. • The observations in Example 3.10 can be formalized as the following result. COROLLARY 3.12. Let H(s) = n(s)/d(s) be a scalar proper rational function. If {A, B, C, D} is a minimal realization of H(s), then kd(s) = a(s) = ani(s\
(3.22)
where a(s) = det (si — A) and a^is) are the characteristic and minimal polynomials of A, respectively, and A: is a real scalar so that kd(s) is a monic polynomial. Proof, The characteristic and minimal polynomials of H{s), pnis), and mnis) are by definition equal to d(s) in the scalar case. Applying Theorem 3.11 proves the result. • Determination via the Hankel matrix There is an alternative way of determining the order of a minimal realization of H(s). This is accomplished via the Hankel matrix, associated with H(s). Given H(s), we express H(s) as a Laurent series expansion to obtain H(s) ^ Ho + H(s) = Ho+ His-^ + H2S~^ + H^^s'^ + • • •,
(3.23)
where H{s) is strictly proper and the real pXm matrices HQ, / / i , . . . are the Markov parameters of the system. They can be determined by the formulas Ho =
limH(s),
Hi =
lims(H(s)-Ho\
H2 = lim s\H(s)
-Ho-
His~'l
and so forth. DEFINITION 3.4. The Hankel matrix MHH, j) of order (/, j) corresponding to the (Markov parameter) sequence H\, H2,... is defined as the ip X jm matrix given by Hi H2 MH(iJ)
H2 H3
Hj+i (3.24)
= Hi
Hi+i
Hi+j-i\
THEOREM 3.13. The order of a minimal realization of H(s) is the rank of Mnir, r), where r is the degree of the least common denominator of the entries of H(s), i.e., r = degmnis). Proof, Let {A, B, C, D] be any realization of H{s) of order n. The Markov parameters satisfy the relationships Hi = CA'-^B,
/ = 1,2,...,
(3.25)
399 CHAPTER 5 :
Realization Theory and Algorithms
400 Linear Systems
(refer to Exercise 2.63 in Chapter 2). Therefore, the Hankel matrix Mnir, r) can be written as C CA
Mnir, r)
[B,AB,...,A'-^Bl
026)
Using Sylvester's Inequahty, we have rank Mnir, r) < n, since the common dimension in the product given by (3.26) is n. This imphes that any reaHzation of H(s) is of order higher than or equal to rank Mnin r). We now must show that there exists a realization of order exactly equal to rank Mnir, r), where r is the degree of the least common denominator of the entries of H{s). Let {A, B, C} be given by 0„
0„
Hi
and
B = H2
A= -dolp
-dil^ Up
where A E RP'^'P^
C = Up, 0,
,0],
^p
^p
-dr-\Ip
B E RP''''^,
Hr and C G RP^'P' with mnis)
=^ s' + dr-is''^
-h ••• -h
d\s + d{), the minimal polynomial of H{s). We have that {A, B, C) is an observable realization of H{s) (see Section 5.4). Furthermore, note that |^/ - A| = (mH(s))P. Now C CA rankMH{r,r) = rank [B,..., A'-^B] = rank [B,..., A'-^B]. This is true beC CA
CA' = Ipr, as can easily be seen from above. Note now that mniA) = 0.
CA'-^ This is true since (m//(A))^ = 0, in view of the Cay ley-Hamilton Theorem. Therefore, A\ i > r, can be expressed as a linear combination of /, A , . . . , A'""^ which implies that rank Mnin r) = rank[B, ...,A'-^B] - rank [ 5 , . . . , AP^'^BI Therefore, the rank of the controllability matrix of the above realization is equal to rank Mnir, r). Hence, if {A, B, C} is reduced by means of a transformation to a standard uncontrollable form
Ai
An
0
Ai.
[Ci, C2] \ (see Subsection 3.4A) with (Ai, Bi) controllable, then
{Ai, Bi, Ci} will be a controllable and observable (minimal) realization of H(s) of order equal to rankMnir, r), since this is the rank of the controllablity matrix of (A, B) and the dimension of Ai. • E X A M P L E 3 . 1 1 . Let
H{s) =
1 s+ 1 -1 {s + l){s + 2)
s+ 1 1 s+2
Here the minimal polynomial is muis) = (s+l)(s + 2), and therefore, r = deg mnis) 2. The Hankel matrix M//(r, r) is then Mnir, r) = MH(2, 2)
Hi
H2
H2
H3
an r/7 X rm = 4 X 4 matrix, and
401 Is s+ 1 s ^+ 2
5+ 1 H\ = \im sH(s) = lim —s S^ 00 s^ °° {s + 1)(^ + 2)
CHAPTERS:
1 2 0 1
and H2 = lim y^(i/(5) 5->oo
r
= lim
His- ')
.2
2s'
-.2
5—»oo
[(s+ l)(s + 2) -s
[(s+ l)(^ + 2) Similarly, H^ =
1 3
s^ s+2 -2s s+ 1 = -2s s + 2.
s+ 1
= lim
o
2 .No> N 4
rank MH (2, 2) = rank
1 0 1 1
2 1 -2 -2
-1 -1 1 3
-2 -2 = 3, 2 4
which is the order of any minimal realization, in view of Theorem 3.12. The reader should verify this result, using Theorem 3.11. • Ills 21s LO - 1 / ^ as m Example 3.9. Here r = degmnis) degmnis) = degs = 1. Now, the Hankel matrix Mnir, r) EXAMPLE 3.12. Consider the transfer function matrix H(s) =
n
2
MH(1 I) = HI = lim^-^oo sH(s) = . Its rank is 2, which is the order of a [0 - 1 minimal realization of H(s). This agrees with the results in Example 3.9.
D. Minimality of Realizations: Discrete-Time Systems All definitions and theorems given thus far in this section for the continuous-time case, apply directly to the discrete-time case with no substantial changes. Thus, minimal or irreducible realizations are defined as in Definitions 3.1 and 3.2, where H(k, i) and H{z) should be used in place of H{t, r) and H(s), respectively. The main criteria for establishing minimality are given by results that are essentially the same as Theorems 3.9 and 3.10. The McMillan degree of H(z) is as defined in Definition 3.3, while results that are essentially the same as Theorems 3.11 and 3.13 provide a means of determining the order of the minimal realizations.
Realization Theory and Algorithms
402 Linear Systems
The fact that the results on minimality of realizations in the discrete-time case ^^^ essentially identical to the corresponding results for the continuous-time case is not surprising since we are concentrating here on the time-invariant cases for which the transfer function matrices have the same forms: H(s) = C(sl - A)~^B + D and H(z) = C(zl - A)~^B + D. Accordingly, the results on how to generate 4-tuples {A, B, C, D] to satisfy these relations are of course the same. The realization algorithms developed in Section 5.4 apply directly to the discrete-time case as well, since they are algorithms for time-invariant realizations of transfer matrices. The differences in the realization theory between continuous- and discrete-time systems arise primarily in the time-varying case (compare Theorems 3.1 and 3.5, for example). However, these differences are rather insignificant.
5.4 REALIZATION ALGORITHMS In this section, algorithms for generating time-invariant state-space realizations of external system descriptions are introduced. In particular, it is assumed that a proper rational matrix H{s) of dimensions p X m is given for which a state-space reahzation {A, B, C, D} of H{s) given by i = Ax + Bu, y = Cx + Du such that C{sl - A)~^B -\- D = H(s) has been derived (see Theorem 3.3). This problem is equivalent to realizing H(t, 0) = 5£~^[H(s)], A brief outline of the contents of this section follows. Realizations of H(s) can often be derived in an easier manner if duality is used, and this is demonstrated first in this section. Realizations of minimal order are both controllable and observable, as was shown in the previous section. To derive a minimal reahzation of H(s), one typically derives a realization that is controllable (observable) and then extracts the part that is also observable (controllable), using the methods of Subsection 3.4A of Chapter 3. This involves in general a two-step procedure. However, in certain cases a minimal realization can be derived in one step, as for example, when H(s) is a scalar transfer function. Algorithms for realizations in a controller/observer form are discussed first. In the interest of clarity, the SISO case is presented separately, thus providing an introduction to the general MIMO case. Realization algorithms, where A is diagonal or in block companion form, are introduced next. Finally, balanced realizations are addressed. It is not difficult to see that the above algorithms can also be used to derive realizations described by equations of the form x(k + 1) = Ax(k) + Bu{k), y{k) = Cx(k) + Du(k) of transfer function matrices H(z) for discrete-time time-invariant systems. Accordingly, the discrete-time case will not be treated separately in this section.
A. Realizations Using Duality If the system described by the equations x = Ax-\- Bu, y = Cx + Du is a realization of if (^), then H{s) = C(sl - AY^B + D.
(4.1)
If H(s) = H^{s), then x = Ax + Bu dind y = Cx + Du, where A =- A^,B = C^, C = B^, and D = D^, is a reahzation of H(s) since in view of (4.1),
A^y^C^
= C(sl -Ay^B
+D^
+ D.
(4.2)
The representation {A, B, C, D] is the dual representation to {A, B, C, D}, and if {A, B, C, D} is controllable (observable), then {A, B, C, D} is observable (controllable) (see Chapter 3). In other words, if a controllable (observable) realization {A, B, C, D} of the p X m transfer function matrix H{s) is known, then an observable (controllable) realization of the m X /> transfer function matrix H{s) = H^(s) can be derived immediately: it is the dual representation, namely, {A, B, C, D} = {A^, C^, B^, D^}. This fact is used to advantage in deriving realizations in the MIMO case, since obtaining first a realization of H^(s) instead of H(s) and then using duality, leads sometimes to simpler, lower order, realizations. Duality is very useful in realizations of symmetric transfer functions, which have the property that H(s) = H^(s), as, e.g., in the case of SISO systems where H(s) is a scalar. Under these conditions, if {A, B, C, D} is a controllable (observable) realization of H(s), then {A^, C^, B^, D^} is an observable (controllable) realization of the same H{s). Note that in this case, H{s) = C(sl - Ay^B
+ D = H^{s) - B^{sl - A^y^C^
-h D^.
In realization algorithms of MIMO systems, a realization that is either controllable or observable is typically obtained first. Next, this realization is reduced to a minimal one by extracting the part of the system that is both controllable and observable, using the methods of Subsection 3.4A. Dual representations may simplify this process considerably. In the following, we summarize the process of deriving minimal realizations for the reader's convenience. Given a proper rational p X m transfer function matrix H(s), with lims-^ooH(s) < 00 (see Theorem 3.3), we consider the strictly proper part H(s) = H(s) - lirris-^a:, H(s) = H(s) - D [noting that working with H(s) instead of H(s) is optional]. 1. If a realization algorithm leading to a controllable reahzation is used, then the following steps are taken, H(s) -^ (H(s) = H^(s)) -^ {A, B, C}-> {A = A^, B = C^,C = B^l
(4.3a)
where {A, B, C] is a controllable realization of H{s) and {A, B, C} is an observable realization of H(s). 2. To obtain a minimal realization. {A, B, C}
0
An A2
[Ci, C2]
CHAPTERS:
Reahzation Theory and Algorithms
H(s) = H^(s) = B^{sl-
403
(4.3b)
where {A, B, C} is an observable realization of H(s) obtained from step (1), and (Ai, Bi) is controllable (derived by using the method of Subsection 3.4A), then {Ai, Bi, Ci} is a controllable and observable, and therefore, a minimal realization of H(s), and furthermore, {Ai, Bi, Ci, D} is a minimal realization of H{s).
404 Linear Systems
B. Realizations in Controller/Observer Form We shall first consider realizations of scalar transfer functions H(s). Single-input/single-output (SISO) systems (p = m = 1) Let H{s) =
n(s) d(s)
bnS^ + • • • + bis + Z?o s^ + an-\s^ 1 + • • • + a\s + UQ
(4.4)
where n{s) and ^(5*) are prime polynomials. This is the general form of a proper transfer function of (McMillan) degree n. Note that if the leading coefficient in the numerator n{s) is zero, i.e., bn = 0, then H(s) is strictly proper. Also, recall that (1) + bou, bnu^""^ + '" + biu^'^
(4.5a)
72-1
d{q)y{t) = {q"" + an-iq"" ' + • • • + ^ i ^ + ao)y{t) = (bnq^ + -"bxq + bo)u(t) = n(q)u(t),
or
(4.5b)
where q = d/dt, the differential operator. This is the corresponding nth-order differential equation that directly gives rise to the map y(s) = H(s)u(s) if the Laplace transform of both sides is taken, assuming that all variables and their derivatives are zero 3tt = 0. Controller form realizations Given n(s) and d(s), we proceed as follows to derive a realization in controller form. 1. Determine C j G R"" and Dc ^ R so that n(s) = CcS(s) + Dcd(sX where S(s) = [l,s,..., implies that
(4.6)
s^~^]^ is an n X 1 vector of polynomials. Equation (4.6) Dc = limH(s)
= bn.
(4.7)
Then n(s) — bnd(s) is in general a polynomial of degree n- I, which shows that a real vector Q that satisfies (4.6) always exists. If bn = 0, i.e., if H(s) is strictly proper, then from (4.6) we obtain Q = [bo,..., bn-i], i.e., Cc consists of the coefficients of the n- I degree numerator. If bn # 0, then (4.6) implies that the entries of Q are a combination of the coefficients bt and a/. In particular, Cc = [bo - bnao, bi - bnUi,..., Z?„_i - Z?«a„-i].
(4.8)
2. A realization of H(s) in controller form is given by the equations
yi'C
-^^C "^ C
JJ Q Id'
0
1
0
0
-flo
y = CcXc + DcU.
0 Xc +
-a„-\ (4.9)
405
The n states of the reaUzation in (4.9) are related by -^/+i
= Xj
or
X, i+i
.(0
n-l
and
-aoxi
n-\
^atXi^i^u i=\
CHAPTER 5 :
n—1,
1, -aoxi
Realization Theory and Algorithms
.(i)
i=l
It can now be shown that xi satisfies the relationship d{q)xi{t)
= u{t),
y{t) =
n{q)xi{t),
(4.10)
where q = d/dt,
the differential operator. In particular, note that d{q)xi{t) Sn) which in -d{q)xi x^\ u(t) because Xn = — E/Lo1...W ^iH ~^ » ^ » In) derived from Xn = x\' > - l )^\ implies that —d{q)xi + w = 0. The view of Xn = -^i"^ relation y{t) = n{q)xi {t) can easily be verified by multiplying both sides ofn{q) = CcS{q) -\-Dcd{q) given in (4.6) by xi. LEMMA 4.1. The representation (4.9) is a minimal realization of H(s) given in (4.4). Proof, We must first show that (4.9) is indeed a realization, i.e., that it satisfies (4.1). This is of course true in view of the Structure Theorem in Subsection 3.4D of Chapter 3. Presently, this will be shown directly, using (4.10). Relation d{q)x\{t) = u{t) implies that x\{s) = (d(s))~^u(s). This yields for the state that x(s) = [xi(s),...,Xn(s)]'^ = [l,s,...,s''~^]'^xi(s) = S(s)(d(s))~^i2(s). However, we also have x{s) = {si — Ac)~^Bcu{s). Therefore, {sI-Ac)S{s)=Bcd{s).
(4.11)
Now CcisI-Ac)-^Bc+Dc =CcS{s){d{s))-^+Dc = (QSis) + Dcd{s)){d{s))-^ = n{s)/d{s) =H{s), i.e., (4.9) is indeed a realization. System (4.9) is of order n, and is therefore, a minimal, controllable, and observable realization. This is because the degree of H{s) is n, which in view of Theorem 3.11, is the order of any minimal realization. Controllability and observability can also be established directly by forming the controllability and observability matrices. The reader is encouraged to pursue this approach. • According to Definition 3.3 given in Section 5.3, the McMillan degree of a rational scalar transfer function H{s) = n{s)/d{s) is n only when n{s) and d{s) are prime polynomials; if they are not, all cancellations must first take place before the degree can be determined. If n{s) and d{s) are not prime, then the above algorithm will yield a realization that is not observable. Notice that realization (4.9) is always controllable, since it is in controller form. This can also be seen directly from the expression
ro [5c,Ac5c,...,A^
Be
0 1
•••
1" (4.12)
X
which is of full rank. The realization (4.9) is observable if and only if the polynomials d{s) and n{s) are prime. In Fig. 5.2 a block realization diagram of the form (4.9) for a second-order transfer function is shown. Note that the state xi{t) and X2{t) are taken to be the voltages at the outputs of the integrators. This is common when realizing transfer functions using analog computer circuits.
406
_b2_
Linear Systems
b-\ — b23-\
u
. ^
X2
.
LU
^'i
*
bo - ^231
+ +
Efi ^
lifi
FIGURE 5.2 Block realization of H(s) in controller form of the system 0 -ao
» =
11
+ 0 w,y = [bo-b2ao,bi - M l ]
-ai\ \x2 1 b2S^ + bis + bo .2 + ais-\-ao
-\-b2U;
Observer form realizations Given the transfer function (4.4), the nth-order realization in observer form is given by XQ ^
/\QXQ
0 1
-\-IJQU
• •
0 0
-ao
bo - bnao bi - bnai
—ai ^0 +
0
• •
1
—Cln-l
bn-l
y = CoXo + DoU = [0,0,..., 0, l]xo + bnU.
-K^n-1
(4.13)
This realization was derived by taking the dual of realization (4.9). Notice that A^ = Al,Bo=Cl,Co=Bl, midDo=Dl. LEMMA 4.2. The representation (4.13) is a minimal realization of H{s) given in (4.4). Proof, Note that the observer form realization {Ao,Bo,Co,Do} described by (4.13) is the dual of the controller form realization {Ac,Bc,Cc,Dc} described by (4.9), used in Lemma 4.1. • The realization (4.13) can also be derived directly from H{s), using defining relations similar to (4.6). In particular. Bo and Do can be determined from the expression n{s)=S{s)Bo^d{s)Do, (4.14) where 5(^) = [1,^,. . . , ^ ^ ~ ^ ] . It can be shown (by taking transposes) that the corresponding relation to (4.11) is now given by S{s){sI-Ao)=d{s)Co (4.15) and that d{q)z{t) = n{q)u{t),y{t) = z{t) corresponds to (4.10).
(4.16)
Figure 5.3 depicts a block realization diagram of the form (4.13) for a secondorder transfer function.
407 CHAPTERS:
Realization Theory and Algorithms
FIGURE 5.3 Block realization of H(s) in observer form of the system Xi 0 -^o]\x{ + bo- - b2ao + b2U\ = u.y = [0,1] 1 -ai\ U2. Pi- - bzai X2_ H{s) =
b2S^ + bx 5 +
EXAMPLE 4.1. We wish to derive a minimal realization for the transfer function H{s) = (s^ + s - l)/(s^ -\- 2s'^ - s - 2). Consider a reahzation {A^ Be, Cc, Dc}, where (Ac, Be) is in controller form. In view of (4.6) to (4.9), Dc = lims^oo H(s) = 1 and n(s) = s^ -\- s — 1 = CcS(s) + Dcd(s), from which we have CcS(s) = (s^ + s — I) (s^ + 2s^ - s -2) = -2s^ + 2^ + 1 = [1, 2, -2][1, s, s^f. Therefore, a realization of H{s) is Xc = AcXc + BcU, y = CcXc + DcU, where Ac =
0 1 0" 0 0 1 2 1 -2
ro Be =
0 1
Cc = [ 1 , 2 , - 2 ] ,
Dc
This is a minimal realization (verify this). Instead of solving n(s) = CcS(s) + Dcd(s) for Cc as was done above, it is possible to derive Q by inspection after H(s) is written as H(s) = H(s) + lim^(^) = H(s) + Dc,
(4.17)
where H(s) is now strictly proper. Notice that if H(s) is given by (4.4), then Dc = bn and H(s)
Cn-lS''
+ ••• + Ci5 + Co
• Un-is^ 1 + ••• + ais + ao
(4.18)
where in fact, c/ = hi - bnUi, i = 0 , . . . , /t - 1. The realization {Ac, Be, Cc) of H(s) has (Ac, Be) precisely the same as before; however, Cc can now be written directly as Cc = [Co, Ci,
...,Cn~ll
(4.19)
i.e., given H(s) there are three ways of determining Q : (i) using formula (4.8), (ii) solving CcS(s) — n(s) — Dcd(s) as in (4.6), and (iii) calculating H(s) = H(s) - lim^^^oo H(s). The reader should verify that in this example, (i) and (iii) yield the same Q = [1, 2, - 2 ] as in method (ii). Suppose now that it is of interest to determine a minimal reahzation {Ao, Bo, Co, Do}, where (Ao, Co) is in observer form. This can be accomplished in ways completely
408 Linear Systems
analogous to the methods used to derive realizations in controller form. Alternatively, one could use duality directly and show that
AQ
— A„ —
0 1 0
0 2 1 0 1 -2
Co = Bl = [0,0,1],
Bo - Cc
-
Do = Di = I
is a minimal realization, where the pair (A^, Co) is in observer form.
•
EXAMPLE 4.2. Consider now the transfer function H(s) = (s^ - l)/(s^ +2s^ - s - 2), where the numerator is n(s) = ^^ - 1 instead of ^-^ + s - 1, as in Example 4.1. We wish to derive a minimal realization of H(s). Using the same procedure as in the previous example, it is not difficult to derive the realization Ac
1 0^ 0 1 1 -2
0 0 2
ro Be =
0 1
Cc = [1,1, - 2 ] ,
Dc
This realization is controllable, since (A^ Be) is in controller form (see Exercise 5.11); however, it is not observable, since rank € = 2 < 3 = n, where 0 denotes the observability matrix given by
r 1 1 -2 = -4 -1 5 QA?. 10 1 -11 Cc '
C-c-^c
(show this). Therefore, the above is not a minimal realization. This has occurred because the numerator and denominator of H(s) are not prime polynomials, i.e., ^ - 1 is a common factor. Thus, strictly speaking, the H(s) given above is not a transfer function, since it is assumed that in a transfer function all cancellations of common factors have taken place. (See also the discussion following Lemma 4.1.) Correspondingly, if the algorithm for deriving an observer form would be applied to the present case, the realization {Ao, Bo, Co, Do] would be an observable realization, but not a controllable one, and would therefore not be a minimal realization. To obtain a minimal realization of the above transfer function H{s), one could either extract the part of the controllable realization {Ac, Be, Cc, Dc) that is also observable, or simply cancel the factor 5 - 1 in H{s) and apply the algorithm again. The former approach of reducing a controllable realization will be illustrated when discussing the MIMO case. The latter approach is perhaps the easiest one to apply in the present case. We have 1
His) =
2^2
s^ + s + \ s^ + 3s + 2
-2s- 1 + 1, s^ + 3s-\-2
and a minimal realization of this is then determined as 0 -2
1 -3.
Be
Cc
[-1,-2],
Dc
The reader should verify this. Multi-Input-Multi-Output (MIMO) Systems (pm > 1) Let Si(p X m) proper rational matrix H(s) be given with lim^^oo H(s) < oo (see Theorem 3.3). We now present alogrithms to obtain realizations {Ac, Be, Cc, Dc} of H(s) in controller form and realizations {Ao, Bo, Co, Do} of H(s) in observer form.
Minimal realizations can then be obtained by separating the observable (controllable) part of the controllable (observable) realization. Controller form realizations Consider a transfer function matrix H{s) = [nij{s)/dij{s)],i = l , . . . , p , 7 = 1,..., m, and let ij (s) denote the (monic) least common denominator of all entries in the jth column of H{s). The ij{s) is the least degree polynomial divisible by all dij {s)J= 1,..., p. Then H{s) can be written as
H{s)=Nis)D-\s),
(4.20)
a ratio of two polynomial matrices, where N{s) = [nij{s)] and D{s) = diag[ii{s),.. .,im{s)]. Note that nij{s)/ij{s) = nij{s)/dij{s) for / = 1, j = 1,..., m. Let dj = deg ij (s) and assume that dj >1. Define
, p, and all
A(^) = diag (y4i
S{s) = block diag
and
(4.21) \ s'lj-'
and note that S{s) ism\n(=
/
E7=i ^j) ^^ polynomial matrix. Write D{s) = DhA{s) + DiS{s)
(4.22)
and note that D^ is the highest column degree coefficient matrix of D{s). Here D{s) is diagonal with monic polynomial entries, and therefore, D/^ = Im- If. for exam[3^2 j^ ^ 2s] pie, D{s) = \ ^ , then the highest column degree coefficient matrix D^ = 2s , and DiS{s) given in (4.22) accounts for the remaining lower column degree terms in D{s), with D^ being a matrix of coefficients. Observe that \Dh\ 7^ 0, and define the mxm and mxn matrices -Dn'Di
(4.23)
respectively. Also, determine Cc and Dc such that N{s)=CcS{s)^DcD{s),
(4.24)
and note that Dc = lim H{s).
(4.25)
We have H{s) = N{s)D-^ (s) = CcS{s)D-^ (s) +Dc with CcS{s)D-^ {s) being strictly proper (show this). Therefore, only Cc needs to be determined from (4.24). A controllable realization of H{s) in controller form is now given by the equations Here Cc and Dc were defined in (4.24) and (4.25), respectively. ^c-^mi
Be — BcByy
(4.26)
409 CHAPTER 5 :
Realization Theory and Algorithms
410
where Ac = block diag [Ai, A 2 . . . , A^] with
Linear Systems j^djXdj ^j
= \'
Id,
0
0
...
0
G R'^J, j =
Be = block diag
I..., m
and Am, Bm were defined in (4.23). Note that if dj = /JLJ, j = 1 , . . . , m, the controllabiHty indices, then (4.26) is precisely the relation (4.63) given in Section 3.4. LEMMA 4.3. The system {Ac, Be, Cc, Dc} is an w( = 2 J.. 1 ^;)-th-order controllable realization of H(s) with (Ac, Be) in controller form. Proof. First, to show that {Ac, Be, Cc, Dc} is a realization of H(s), we note that in view of the Structure Theorem given in Subsection 3.4D, we have Cdsl - Ae)~^Be + Dc = N(s)D(s)-\whQYQ
D(s) ^
B;,'[A(S)
- AmS(s)l
N(s) ^ CeS(s) + DcD(s).
However, 5(5) = D(s)mdN(s) = A^(5),in view of (4.22) to (4.24). Therefore, Q ( 5 / AcT^Bc + De = N{s)D-\s) = //(5), in view of (4.20). It is now shown that (Ac, Be) is controllable. We write [si
Ac, Be] = [sl -Ac-
BcAm, BcB^] I 0
(4.27)
= [si - Ac, Be]
Brr
and notice that rank [sjl - Ac, Be] = n for any complex Sj. This is so because of the special form of A^, Be. (This is, in fact, the Brunovski canonical form.) Now since \Bm\ ^ 0, Sylvester's Rank Inequality implies that rank [sjl - Ac, Be] = n for any complex Sj, which in view of Subsection 3.4B, implies that {Ac, Be) is controllable. In addition, since Bm = Im^ it follows that (Ac, Be) is of the form (4.69) given in Subsection 3.4D. With dj = )LiJ, the pair (Ac, 5c) is in controller form. • An alternative way of determining Q is to first write H(s) in the form H(s) = H(s) + lim H(s) = H(s) + Dc
(4.28)
where H(s) = H(s) - Dc is strictly proper. Now applying the above algorithm to H(s), one obtains H(s) = N(s)D~^(s), where D(s) is precisely equal to the expression given in (4.20). We note, however, that N(s) is different. In fact, N(s) = N(s) - DcD(s). The matrix Q is now found to be of the form A^(^) -
CcS(sl
(4.29)
Note that this is a generalization of the scalar case discussed in Example 4.1 [see (4.17) to (4.19)]. In the above algorithm the assumption that dj > 1 for all j = 1 , . . . , m, was made. If for some 7, dj = 0, this would mean that the jth column of H(s) will be a real m X 1 vector that will be equal to the jth column of Dc [recall that Dc =
lim^-^oo H(s)]. The strictly proper H(s) in (4.28) will then have its jth column equal to zero, and this zero column can be generated by a realization where theyth column of Be is set to zero. Therefore, the zero column (thejth column) of H{s) is ignored in this case and the algorithm is applied to obtain a controllable realization. A zero column is then added to Be. (See Example 4.4.) Finally, we note that given H(s) = N(s)D~^{s) with A^(^), D(s) not necessarily from (4.20), the above algorithm leads to a controllable realization if dj = column degrees of D(s), provided that D^, the highest column degree coefficient matrix is nonsingular (\Dh\ # 0). This, in fact, means that D(s) is column proper (see Subsection 7.2B). The resulting pair (Ac, Be) is in controllable, companion form but not necessarily in controller form, since Bm = D^^ is not necessarily upper triangular with ones on the diagonal. (See Example 4.5.) Observer form realizations These realizations are dual to the controller form realizations and can be obtained by duality arguments [see (4.2) and Example 4.3]. In the following, observer form realizations are obtained directly for completeness of exposition. We consider the transfer function matrix H{s) = [nij{s)ldij{s)\ i = 1 , . . . , p, j = 1 , . . . , m, and let £i{s) be the (monic) least common denominator of all entries in the /th row of H(s). Then H(s) can be written as H(s) = D'\s)N(s\ where N(s) = [nij(s)] and D(s) = diag [ti{s\ . ..Jp{s)l nij(s)/dij(s) for 7 = 1 , . . . , m, and all / = 1 , . . . , p. Let di = deg (((s), assume that J^- > 1, define
(4.30) Note that nij{s)l'ii{s) =
A(^) = diag ( / s . . . , / ^ ) , S(s) = block diag ([I, ^ , . . . , / ~ ^ ] , / = 1, (4.31) and note that S(s) is a p X n ( = Xf= i di) polynomial matrix. Now, write D(s) = Ms)Dh + S(s)De
(4.32)
and note that Dh is the highest row degree coefficient matrix of D(s). Note that D(s) is diagonal, with entries monic polynomials, so that Dh = Ip, the pX p identity ma\3s^ + 1 2^1 trix. If, for example, D(s) = , then the highest row degree coefficient 2s matrix is Du
and S(s)D^ in (4.32) accounts for the remaining lower row
degree terms of D(s), with D^ a matrix of coefficients. Observe that \Dh\ y^ 0; in fact Dh = Ip. Define the p X p and nX p matrices Cp = Dh^
and
Ap = -DeDh\
(4.33)
respectively. Also, determine Bo and Do such that N(s) = S(s)Bo + D(s)Do.
(4.34)
Do = lim H(s\
(4.35)
Note that and therefore, only Bo needs to be determined from (4.34).
411 CHAPTERS:
Realization Theory and Algorithms
412
An observable realization of H(s) in observer form is now given by
Linear Systems
Xo
= AQXO
+ BQU,
y
— CQXO +
DQU,
where Bo and Do were defined in (4.34) and (4.35), respectively, and (4.36) where Ao = block diag [Ai, A2,..., Ap\ with 0 ... Ai
0
0 0
=
j^diXdi
^di-\
0
Co = block diag ([0,.. .,0,1] E R^^^\ i = 1,..., p), and Ap, Cp is defined in (4.33). Note that (4.36) is exactly relation (4.85) of Section 3.4 if di = vt, i = 1,...,/?, the observability indices. LEMMA 4.4. The system {Ao, Bo, Co, Do} is an n (= 2f=i
S^ + I S + I
. We wish to derive a minimal realization
for H(s). To this end we consider realizations {Ac, Be, Cc, Dc), where {Ac, Be) is in controller form. Here €1(5) = s^, €2(5) = s^, and H{s) can therefore be written in the form (4.20) as H{s) = N{s)D~\s) = {s^ -^\,s+ 1] 0
1 ^ 0 0 0 . Note that n = 0 0 I s s^ di + d2 = 5, and therefore, the realization will be of order 5. Write D{s) = DhA(s)-\Here di = 2,d2 = 3 and A(s)
, S(s)
ro 0 0 0 01
DeS(s), and note that Dh = h, D^ = 0 0 0 0 0 . Therefore, in view of (4.23), Bm —
1 0 0 1
and
Am = -DP
=
0 0
0 0 0 0
0 0 0 0
Here Dc = lims^oo H(s) = [1, 0] and (4.24) implies that CcS(s) = N(s) - DcD(s) = [s^ + 1, 5 + 1] - [s^, 0] - [l,s + 1] from which we have Q = [1, 0, 1, 1, 0]. A controllable realization in controller form is therefore given by x - AcXc + BcU and y = ^~^ C
C
-^-^C
9
413
where
CHAPTERS:
0
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
and
I
0" 0
0 0 0
0 0 1
Be
Ac =
Cc
'0 1
[1, 0,1,1, 0],
= [1,0].
Note that the characteristic (pole) polynomial of H{s) is s'^ and the McMillan degree of H{s) is 3. The order of any minimal reaUzation of H{s) is therefore 3 (see Theorem 3.11). This implies that the controllable fifth-order realization derived above cannot be observable [verify that (A^, Cc) is not observable]. To derive a minimal realization, the observable part of the system [Ac, Be, Cc, Dc) needs to be extracted, using the method described in Subsection 3.4A. In particular, a transformation matrix P needs to be determined so that A = PAcP
Ai
0
A21
A2,
-1 _
and
C = CcP~
where (Ai, Ci) is observable. If B = PBc
[Ci,0],
, then {Ai, 5i, Ci, Di} is a minimal
realization of H{s). To reduce (Ac, Cc) to such standard form for unobservable systems, we let AD = AJ, BD = C j , and Co - B^ and we reduce (AD, BD) to a standard form for uncontrollable systems. Here the controllability matrix is 1 0 0 0 0 1 0 0 1 0 0 0 1 1 0 0 0 1 1 0
0 0 0 0 0
Pj)^ are taken to be the Note that rank ^ ^ = 3. Now if the first three columns of QD first three linearly independent columns of ^ D , while the rest are chosen so that \QD\ ¥= 0 (see Subsection 3.4A), then
and
QD
1 0 0 1 = 1 0 1 1 0 1
QD'-
0 0 0 0 1
—; -]
0 0 0 0 0 1 0 0 1 0 0 0 0 1 0
1 0 0 0 0 1 0 0 -1 - 1
0 0 1 0 0
Realization Theory and Algorithms
414
This implies that
Linear Systems
AD = QDAUQD
=
ADI
ADU
[ 0
— QD ^D —
0
0
1
1
0
0
0
0
1
0
0
-1
0
0
0
-1
1
0
0
0
-1
1
-1 1
AD2
1 •
'
BD
0
0 0
BDI
CD — CDQD
_BDI\
-
0 0
1 0 0 1 1 0
0 0
0 0 Then
A =
Ai
0
.A21
A2.
B = Cl =
^
AT
"0 1 0
0" 1 1
0 0
0 0
^
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
-1
-1
1
1
-1
1
1
C = Bl = [Ci, 0] = [1,0, 0,: 0,0].
Clearly, A = A^, C = B j is in standard form. Therefore, a controllable and observable realization, which is a minimal realization, is given by Xco = Aco^co + BcoU and y = Cco^co + J^coU, where 0 0 0
1 0" 0 1 0 Oj
Bco -
"0 0 1 1 .0 1
Ceo = [1,0,0],
Dc
[1,0].
A minimal realization could also have been derived directly in the present case if a realization {A^, Bo, Co, Do} of H(s), where (Ao, Bo) is in observer form, had been considered first, as is shown next. Notice that the McMillan degree of H(s) is 3, and therefore, any realization of order higher than 3 will not be minimal. Here, however, the degree of the least common denominator of the (only) row is 3, and therefore, it is known in advance that the realization in observer form, which is of order three, will be minimal. A realization {Ao, Bo, Co, Do} of H(s) in observer form can also be derived by considering H^(s) and deriving a realization in controller form. Presently, {Ao, Bo, Co, Do}
is derived directly. In particular, we write H(s) = JD ^(s)N(s) = (s^) ^[s(s^ + 1), s + 1]. Then di = 3[= deg \{s) = deg s \ and A ^ = s^,S{s) = [l,s,s^l Then D(s) = s^ = Ms)Dh + S(s)Di implies that Dh = 1 and De = [0, 0, 0]^. In view of (4.33), we have Cp = 1,
Ap = [0,0, 0 ] ^
Do = liiRs^^ H(s) = [1,0], and (4.34) implies that S(s)Bo = N(s) - D(s)Do = 0 1 0 [s(s^ + 1), 5 + 1] - [s^, 0] = [s,s -\- 1], from which we have Bo . An
[1
1 OJ
observable realization of H(s) is the system x = AoXo + BoU, y = CoXo + DoU, where Ao =
0 0 01 1 0 0 0 1 0
ro r
Bo = 1 1 0 0
Co = [0,0,1],
Do = [1,0]
with (Ao, Co) in observer form (see Lemma 4.4). This realization is minimal since it is of order 3, which is the McMillan degree of H(s). (The reader should verify this.) Note how much easier it was to derive a minimal realization, using the second approach. •
EXAMPLE 4.4. Let H(s)
Here €1(5) = s(s+I)
=
s+1 1 s
. We wish to derive a minimal realization.
with di = 2and€2(^) = lwith(i2 = 0. In view of the discussion
following Lemma 4.3, we let Dc = lims^o,H(s)
We now consider the transfer function H(s) ization.
[0 0
2 s+ 1 1
1 and H(s) 0
2 ] s s+ 1 and determine a minimal real 1
Note that the McMillan degree of H(s) is 2, and therefore, any realization of order 2 will be minimal. Minimal realizations are now derived using two alternative approaches: 1. Via a controller form realization. Here £\(s) = s(s + 1), Ji = 2 , and H(s) = " 2s [s(s + l)]-i = N(s)D-\s). Then A(s) = s^ and S(s) = [hsf, D(s) 5+1
DhA(s) + D^S(s). Therefore, B^r, s(s + 1) - Is^ + [0, 1][1,5 " 2s 1 0 2 , which follows from N(s) -[0,1]. Also, Cc .5+ 1 1 1 CcS(s). Then a minimal realization for H(s) is Ac = 0 1
ro 0
1 and Am =
ro 21 rr 1 1 [s
1] ro] - i j ' Be = .ij' Cc
2 Adding a zero column to 5^, a minimal realization of H(s) is now derived as 1 A =
B =
0 1
0 0
c=
0 1
2 1
D =
0 0
1 0
We ask the reader to verify that by adding a zero column to B^ controllability is preserved. 2. Via an observer form realization. We consider H (s) = [2/(s + 1 ) , l/s] and derive a realization in controller form. In particular, ^i = s + I, £2 = s, H^(s) =
415 CHAPTERS:
Realization Theory and Algorithms
416 Linear Systems
s+l 0 5+10^ 0 s 1 0 1 0 [2,1] 0 1 1 0" 0 0
0 s
[2,1]
1 0 s 0 , and S(s) = . Then D(s) = 0 1 0 s 0 1 0 = DhA{s) + DiS(s)mdB^ = 0 1 1
,di = d2 = I, A(s) =
1 0
0 s 1 0
0 s
0 1 0 0
1 0
0 Also, Cc = [2, 1], from which we obtain A^(^) = [2, 1] = 0 = CcS(s). Therefore, a minimal realization {A, B, C} of H^(s) is 1 0
0 , [2, 1] [. The dual of this is a minimal realization of H(s), 1 -1 0 1 0 . Therefore, a minimal realiza, and Co namely, Ao = ,Bo 0 0 0 1 tion of H(s) is r-1 0
0 0
EXAMPLE 4.5. Let 1 ^2 + 1 H(s) =
2 1
B =
1 ^2+1 0
0 0
\l \s
I
1 0
C =
s-l 0
0 1
D =
s+ 1 0 S 5^ + 1
0 0
1 0
N(s)D'\s).
^ s+ 1 We wish to derive a minimal realization based on the given factorization N(s), D{s). Since the McMillan degree of H(s) is 3 (show this), any realization of order 3 will be minimal. Note that in the present case di = I and J2 = 2, the degrees of the first and second columns of D(s), respectively. Then A(^)
0 ,S(s) s'
0
1 0
0 1
0 , and s
1 0 1 0 0 1 0 01 0 + 0 1 0 0 1 ^J = DhA(s) + D^S(s). Here \Dh\ # 0, 1 1 0 and therefore, the algorithm above still applies. Note that Dh is not in upper triangular form with ones on the diagonal, and therefore, the resulting controllable realization will not be in controller form; A^ however, will be in companion form. Let B^ = Dj^^ = 0 0" -1 1 0 -D-,'D, and ,A, 1 0. 1 -1 1 ri 0" 1 - 1 11 "0 01 s+l 0 0 1 + N(s) = CcS(s) + DcD{s) = s s^ +I .-1 0 OJ .1 OJ D(s) =
LO
S_
A minimal realization is now given by 0 Ac =
0 ,
Be -
0
1
0"
0 -1
0 1
-1 Cc
Verify this.
-1 r 0
0
Dc =
0 1
0 0
417
C. Realizations with Matrix A Diagonal
CHAPTER 5 :
When the roots of the minimal polynomial mnis) of H(s) are distinct, there is a realization algorithm due to Gilbert [2] that provides a minimal realization of H(s) with A diagonal. Let mnis) = / + dr-is'' '-{-'"+dis
-^ do
(4.37)
be the (monic) least common denominator of all nonzero entries of the p X m matrix H(s) which in view of Section 3.5, is the minimal polynomial of H(s). We assume that its r roots A/ are distinct and we write (4.38) i=i
Note that the pole polynomial of H(s\ PH(S), will have repeated roots (poles) if PH(S) ¥= mnis) (see Section 3.5 and Example 4.6). We now consider the strictly proper matrix H(s) = H(s) - lims-^oo H(s) = H(s) - D and expand it into partial fractions to obtain H(s)
1 N(s) rriHis)
±7^^.''-
(4.39)
!=1
The pX m residue matrices R, can be found from the relation Ri = lim(s - \i)H(s).
(4.40)
We write Ri = CiBi,
/ - l,...,r,
(4.41)
where C/ is a /? X pi and Bi is a p/ X m matrix with p/ = rank Ri < min (p, m). Note that the above expression is always possible. Indeed, there is a systematic procedure of generating it, namely, by obtaining an LU decomposition of Ri (refer to the Appendix). Then All Upi A2/;P2
)
B=
Bi B2
(4.42)
Br
^rlpr.
D = lim/ i(s)
C = [Ci,C2,...,C,],
is a minimal realization of order n = X L I Pi LEMMA 4.5. Representation (4.42) is a minimal realization of H(s). Proof, It can be verified directly that C(sl - A)~^B + D = H(s), i.e., that (4.42) is a realization of H(s).To verify controllability, we write
= [B^AB,
'B] =
X\Im
B2
Br
Xrlm> '••
Aj
Im
Realization Theory and Algorithms
418 Linear Systems
The second matrix in the product is a block Vandermonde matrix of dimensions mr X mn. It can be shown that this matrix has full rank mr since all A/ are assumed to be distinct. Also note that the (n = %pi) X mr matrix block diag [Bi\ has rank equal to 2/^=1 rank Bt = 2[=i P/ = n ^ mr. Now, in view of Sylvester's Rank Inequality, as applied to the above matrix product, we have n + mr - mr ^ rank ^ < min (n, mr), from which rank % = n. Therefore, {A, B, C, D} is controllable. Observability is shown in a similar way. Therefore, representation (4.42) is minimal. • EXAMPLE 4.6. Let
H(s) = Is+l
s(s+l)i
Herem/zW = ^(5+1) with roots Ai = 0, A2 = - 1 distinct. We write//(5) = (l/s)Ri + [l/(s + l)]/?2, where Ri = lims^o sH(s) = lim^^o
0 2
\ims^-i(s-\-l)H(s) = lim^L 5 - » - l
1 2s s+ 1
0 1 s+l\
n 01 — 0
1
.Ri
0 , pi = rank Ri = 2, and p2 = -1
rank R2 = I, i.e., the order of a minimal realization is n = pi + p2 = 3. We now write Ri =
1 0 0 1
Ri
0 2
1 0 1 0 = C,B, 0 1 0 1 [2 - 1] = C2B2.
-
Then A=
C
A1/2 0
01 _ A2J
[Ci, C2] =
0
0 0
01 0
[0
0
-ij
ro
B =
«1
kJ
n = 0 .2
01 1 -1.
ri 0 01 [0
1
1
is a minimal realization with A diagonal (show this). Note that the characteristic polynomial of H{s) is PH{S) = s'^(s +1), and therefore, the McMillan degree, which is equal to the order of any minimal realization, is 3, as expected. •
D. Realizations with Matrix A in Block Companion Form The realizations derived using the algorithms described below are in general either controllable or observable and of order mr or pr, where r is the degree of the minimal polynomial of ^(.s*). Most often it is also necessary to use the methods of Subsection 3.4A to reduce these realizations to the standard forms for unobservable or uncontrollable systems and in this way derive minimal realizations.
Using the numerator polynomial matrix
419
Consider a proper pxm matrix H{s) and let
CHAPTER 5 :
mnis)
-dr-lS
r-l
Realization Theory and Algorithms
-dis^do
be its minimal polynomial, as in (4.37). The polynomial mnis) is the monic least common denominator of all entries of H{s). Let Nh{s) = mH{s)H{s) be a polynomial matrix and write (4.43)
N}y{s) = CbcSb{s) ^Dbcmnis),
where St{s) = [Im^slm^ ...^s^~^Im]. Note that Dj^c = lim^^ooH(^), and therefore, C}jc is the only unknown in (4.43). A solution Q^ always exists since the highest possible degree in N}y{s) — Df^c^nis) is r — 1 because H{s) is proper. Expression (4.43) is analogous to (4.24). In addition, we can also write mnis) = s^ — r-UT , from which we have -dQ,...,-dr-i][\,s, ..,y mH{s)Im
= S^im — [—dolm^ • • • ,
—dr-lIm\Sb{s)^
(4.44)
which corresponds to D{s) in (4.22). A controllable realization in block controllable companion form is given by x -Ai,cX + Bi,cU^y = ChcS -\-DfycU with Q^ and Df^c specified in (4.43) and 0.
0„
o„
B,be
A be
-dolm
-d\Im
•
(4.45)
-dr-]Irl^m
Note that if the strictly proper matrix H{s) = H{s) — lim^^ooH(^) = H{s) — D^c is used, then Nt{s) ^ mH{s)H{s) = Rr-is'-^ + • • • +/^o, where /^/, / = 0 , . . . , r — 1, are real pxm coefficient matrices. It is now not difficult to see that in this case Q , = [/^o,/^i,...,/^r-i]. (4.46) The realization {A/^c? ^/?C5 Qc? ^/?c} in (4.45) is of order mr and is a direct generalization of the SISO realization (4.9) to the MIMO case. In general it is only controllable but not observable. It is also observable when m = 1, and therefore, it is minimal for m = 1. This is true because of Theorem 3.11 and the fact that in this case r = deg mH{s) = deg PH{S),
the McMillan degree of
H{s).
L E M M A 4.6. The representation (4.45) is a controllable realization of H(s). Proof, The proof is similar to the proof of Lemma 4.1. First to show that (4.45) is a realization, we consider the rmxm matrix X{s) = [X[ (s),. ..,Xj^{s)]'^ = {si — Ai,c)~^Bi,c. Then sX —A^cX = B^c and sXi = Xi^i,/ = 1,...,r — 1, or X^+i = s^Xi,i= 1,...,r — 1. Also, sXf — [—dQlm,---,—dr-\Im\X = Im, which implies, in view of X^+i = s^Xi, that mH{s)X\ = I^. Thus, Xi = {mH{s))~^s^~^Im,i = 1,... ,r, and in view of X{s) = {sI-Abc)-^Bbc, {sI-Abc)Sb{s)=BbcmH{s), (4.47)
420 Linear Systems
which is of course analogous to (4.11). (Refer also to the algorithm for the controller form realization in the MIMO case and the Structure Theorem in Subsection 3.4D, where a similar relation is valid.) Note that \sl — A^J = (mnis))^ (show this). To show that (Ab^, Bbc) is controllable, we observe that Om
[Bbc, AbcBbc,
He Bbc
.,Air'Bbc]
0,
= X
X
(4.48) which has full rank mr, since the first mr columns are linearly independent. We may also easily obtain an observable realization of H{s) using duality. In particular On
-dolp -d\Ip
0, 0„
Ri Bbo
Abo =
(4.49) Rr - l
-dr-llpj
In
LOP
Dbo = lim H(s)
Cbo = [0/7.
is an observable realization of order pr (use duality arguments to prove this). In both of the above cases of controllable realization {Abo Bbc, Cbo Dbc} or observable realization {Abo, Bbo, Cbo, Dbo}, the methods of Subsection 3.4A may be used to obtain minimal realizations [see (4.3b)]. EXAMPLE 4.7. Let H(s) .5+1 5(5-+ 1). as in Example 4.6. We wish to determine an observable realization with A in block companion form. To this end, duality will be used and the procedure described by (4.3a) will be followed. In this case we have mnis) = s^ + s = s^ + dis + do, from which we conclude that r = 2, Ji = 1, and Jo = 0. Note that H(s) is stricdy proper. Let H(s) = H^(s) = 1 s + 1 2s = Nb(s)(mH(s)) ^ and write 0 1 s(s + 1) 1 0
Nb(s) =
2 1 0 s+ 0
0 = Ris + /?o. 1
(4.50)
Then a controllable realization of H(s) is given by
Ahr
=
O2
-doh
0
0
1
0
0
0
0
1
h
Bbc —
-dih.
0
0
-1
0
0
0
0
-1
[Ol
[h\
"0 0
0" 0
1 0
0 1
421
1 0 : 1 2
Cbc — [^0> ^ l ]
0
CHAPTERS:
1 : 0 0
Realization Theory and Algorithms
and an observable realization of H(s) is given by
Abo
0
0
:
0
0
0
0
:
0
0
"1 0
0" 1
1 _2
0 0
Bbo — Cz?c
— ^Z?c
Cbo - Bi ^bc —
1
0
:
-1
0
0
1
:
0
-1
0 0
1
0
0 0
0
1
Note that the system {Abo^ Bboy Cbo) is not controllable. We ask the reader to use the procedure shown in (4.3b) to obtain a minimal realization. The order of any minimal realization is 3 (why?). • Using the Hankel Matrix We consider a proper p X m transfer function matrix H(s) and let muis) = s^ + dr-is^
r-l
+ • • • + dis + do
be its minimal polynomial, as in (4.37). We write (4.51)
H{s) = HQ+ His-^ + H2S~^ + • • •, where the Hi are the Markov parameters of the system. Let Op
Ip
.. •
Op -dolp
0, -dilp
.. • .. •
0,
A =
B =
C = [lp,{)p,...,Opl
ip -~dr-llp_
Hi H2 (4.52) Hr
D = Ho.
LEMMA 4.7. The representation (4.52) is an observable realization of H(s). Proof, To show that (4.52) is a realization, we note that {A, B, C, D} is a realization of H{s) if and only if D = HQ and CA^~^B = ///, / = 1, 2 , . . . (prove this, referring to Exercise 2.63). Here C CA
(4.53)
ICA'-^ and therefore, CA^'^B = Hi, i = 1 , . . . , r. To show that this is also true for / = r + 1 , . . . , a relationship between Hi and di is required. In particular, we let H{s) = H{s) - HQ and we write
422
+ '" +dis + doXHis ^ +H2S ^ + • • •)
mH(s)H(s) = (s' + dr-is'
Linear Systems
(4.54)
'+Ro Equating coefficients of equal powers of s, we obtain .r-2
.
s '.
and
Hi
Rr-ly
H2
Rr-2 ~
dr-\H\,
Hr = RQ ~ dr-iHf-i
Hr+i = —dr-\Hr+i-\
(4.55) — • • • — d\Hiy
— ••• — d{)Hi,
i = 1, 2, ,
Using the above relations, it can be shown that CA^~^B = Hi fori = 1, 2, Now, in view of (4.53), the observability matrix has rank pr, which is equal to the order of the system. Therefore, the system is observable. • We can now use duality arguments to show that 'Om
••
^m
A =
%
-dolm -dllm
. Om . On,
.. •
,
B =
(4.56)
—dr-\Im_
Im
c = m,H2, ....Hrl
D =-H^
is a controllable realization. EXAMPLE 4.8. Let
- 1 H(s) =
0
5'
1
2
L^ + l ^(5 + 1) as in Example 4.7. We wish to determine an observable realization. Here muis) = s^ + s = 5'^ + (ii5 + (io.fromwhichr = 2, Ji = l,anddo - 0. Let//(5) = /fo+^i^~^ + H^s'^ + • • •. The Markov parameters can be found directly from //Q = lim^y^oo H{s) = 0, Hi 0 -2 ample diHi
= lims-.^s(H(s) -Ho) 0 and so forth, or from 1 4.7 (after the transpose 1 0 n 0 = 0 1 [2 0.
1,2,...; that is,
^ ^^ H2 = lims-..o s\H(s) - (Ho + His-')) = 2 0 (4.55) since /?o, • • • > ^r-i are already known from Ex is taken). We have Hi = Ri = ,H2 — RQ — [2 0 0 0 , and H2+i -diH2+i - doHi = —Hi+i, i = -2 1
ro 0 = -H2 = H3 = -H4 = Hs = •••. Therefore, an observable [2 - 1 .
reahzation (4.52) is given by
A =
O2
h
-doh
-dih.
0
0
1
0
0
0 B =
0
0
-1
0
0
0
0
-1
\Hi] [H2\
1 2
0" 0
0 -2
0 1
C = [/2,02] =
1 0
0
0
0 1
0
0
423
ro 01 D= 0 0
CHAPTERS:
This realization is not controllable, since a minimal realization is of order 3. We ask the reader to use Theorem 3.13 to show this. •
E. Realizations Using Singular-Value Decomposition Internally balanced realizations Given a proper p X m matrix H(s), we let r denote the degree of its minimal polynomial mnis), we write
H(s) = Ho^His-^
+//:2^-^ + . . .
to obtain the Markov parameters Hi, and we define \Hi
...
Hr
Ht
H,r + 1
f' A
T ^ Mnir, r) H 2r-\
Hr
(4.57) H,r+l
H2r
where Mnir, r) is the Hankel matrix (see Definition 3.4) and T, f are real matrices of dimension rp X rm. Using singular-value decomposition (see the Appendix), we write T = K
s 0 0
(4.58)
0
where X = diag [Ai,..., A„] £ R"^" with n = rank T = rank Muir, r), which in view of Theorem 3.13 is the order of a minimal reahzation of H(s). The A, with Ai s A2 ^ • • • > A„ > 0 are the singular values of T, i.e., the nonzero eigenvalues of T^T. Furthermore, KK^ = K^K = Ipr and LL^ = L^L = I,nr- We write T = K,XU
(4.59)
=[K,X"^[X"^LA=VU,
where Ki denotes the first n columns of K, L\ denotes the first n rows of L, K\K\ = In, and LiL[ = /„. Also, V G R'P'''' and U E R^''''^. We let V^ and Lf^ denote pseudoinverses of V and U, respectively (see Appendix), i.e., V
= ^-"^Kl
and
(4.60)
^^+ = - L [^^^-"^ U^ S-
where V^V = In and UU^ = In. Now define
A =
V^fu^,
B = UIL
^p,pr
V,
D = Ho.
(4.61)
where 4,^ = Uh ^e-kl k < ^, i.e., Ik^^ is a fc X € matrix with its first k columns determining an identity matrix and the remaining (- k columns being equal to zero. Thus, B is defined as the first m columns of C/, and C is defined as the first/? rows of V. Note that A G 7?"X^ B G Z^^^^'", C G T^^^", and D G RP'''^, LEMMA 4.8. The representation (4.61) is a minimal reahzation of H(s).
Realization Theory and Algorithms
424 Linear Systems
Proof, It can be shown that CA^'^B = Hi, i = 1, 2,... (see also the proof of Lemma 4.7). Thus, {A, B, C, D} is a reahzation. We note that V and U are the observabiHty and controllabihty matrices, respectively, and that both are of full rank n. Therefore, the reahzation is minimal. Furthermore, we notice that V'^V = UU^ = 2 . Realizations of this type are called internally balanced realizations. • The term internally balanced emphasizes the fact that realizations of this type are "as much controllable as they are observable," since their controllability and observability Gramians are equal and diagonal (see Exercise 5.20). Using such representations, it is possible to construct reasonable reduced-order models of systems by deleting that part of the state space that is "least controllable" and therefore "least observable" in accordance with some criterion. In fact, the realization procedure described can be used to obtain a reduced-order model for a given system. Specifically, if the system is to be approximated by a ^-dimensional model with q < n, then the reduced-order model can be obtained from T = Kqdiag[Xi,...,\q\Lq,
(4.62)
where Kq denotes the first q columns of K in (4.58), and Lq denotes the first q rows ofL. 5.5 SUMMARY The theory of state-space realizations of input-output descriptions given by impulse responses, and the time-invariant case by transfer function descriptions, were studied in this chapter. In Section 5.2 the problem of state-space realizations of input-output descriptions was defined and the existence of such realizations was addressed. In Subsection 5.3A time-varying and time-invariant continuous-time and discrete-time systems were considered. Subsequently, the focus was on time-invariant systems and transfer function matrix descriptions H{s). The minimality of realizations of H{s) was studied in Subsection 5.3C, culminating in two results, Theorem 3.9 and Theorem 3.10, where it was first shown that a realization is minimal if and only if it is controllable and observable, and next, that if a realization is minimal, all other minimal realizations of a given H{s) can be found via similarity transformations. In Subsection 5.3C it was shown how to determine the order of minimal realizations directly from H{s). Several realization algorithms were presented in Section 5.4, and the role of duality was emphasized in Section 5.4A. 5.6 NOTES A clear understanding of the relationship between external and internal descriptions of systems is one of the principal contributions of systems theory. This topic was developed in the early sixties with original contributions by Gilbert [2] and Kalman [4]. The role of controllability and observability in minimal realizations is due to Kalman [4]. See also Kalman, Falb, and Arbib [5].
The first realization method for MIMO systems is attributed to Gilbert [2]. It was developed for systems where the matrix A can be taken to be diagonal. This method is presented in this chapter. For extensive historical comments concerning this topic, see Kailath [3]. Additional information concerning realizations for the time-varying case can be found, for example, in Brockett [1], Silverman [9], Kamen [6], Rugh [8], and the literature cited in these references. Balanced realizations were introduced in Moore [7].
5.7 REFERENCES 1. R. W. Brockett, Finite Dimensional Linear Systems, Wiley, New York, 1970. 2. E. Gilbert, "Controllability and Observability in Multivariable Control Systems," SIAM J. Control, Vol. 1, pp. 128-151, 1963. 3. T Kailath, Linear Systems, Prentice Hall, Englewood Cliffs, NJ, 1980. 4. R. E. Kalman, "Mathematical Description of Linear Systems," SIAM J. Control, Vol. 1, pp. 152-192, 1963. 5. R. E. Kalman, R L. Falb, and M. A. Arbib, Topics in Mathematical System Theory, McGraw-Hill, New York, 1969. 6. E. W. Kamen,"New Results in Realization Theory for Linear Time-Varying Analytic Systems," IEEE Trans, on Automatic Control, Vol. AC-24, pp. 866-877, 1979. 7. B.C. Moore,"Principal Component Analysis in Linear Systems: Controllability, Observability and Model Reduction," IEEE Trans, on Automatic Control, Vol. AC-26, pp. 17-32, 1981. 8. W. J. Rugh, Linear System Theory, Second Edition, Prentice-Hall, England Cliffs, NJ, 1996. 9. L. M. Silverman, "Realization of Linear Dynamical Systems," IEEE Trans, on Automatic Control, Vol. AC-16, pp. 554-567, 1971.
5.8 EXERCISES 5.1. Given the transfer function matrix H{s)
s- 1 s 0
s-2 s+2 5+ 1
0
determine the McMillan degree of H{s) and find a minimal realization for H{s). Verify your results. 5.2. Consider the transfer function matrix H{s)
\s- 1 s+\ 1
1 s^ - \ 0
(a) Determine the pole polynomial and the McMillan degree of II{s), using both the Smith-McMillan form and the Hankel matrix.
425 CHAPTER 5:
Realization Theory and Algorithms
426 Linear Systems
(b) Determine an observable realization of H(s). (c) Determine a minimal realization of H(s). 1 1 s + 1 s^ 1
Hint: Obtain realizations for
(s + 1)(-^ + 5) (s - l)(s^ -9)' for H(s) a minimal realization in controller form.
5.3. Consider the transfer function matrix H(s)
s s-1
and determine
5.4. Consider the transfer function matrix -3s^
His) = I
-6s-2
(^ + 1)^ s
(S + 1)3
3^- 1 (s - 2)(s + 1)3 s (s-2)(s+ 1)3
1 (s - 2)(s + 1)2 s ( ^ - 2 ) ( ^ + 1)2 J
(a) Determine the pole polynomial of H(s) and the McMillan degree of H(s). (b) Determine a minimal realization of H(s) in observer form. 5.5. Consider the scalar proper rational transfer function//(>s')[})(5') = //(5')M(^)], and assume -. Note that this can be done when the that H(s) can be written as H(s) = 2 L i s - Ai
poles A/, / = 1 , . . . , w, are distinct (see Subsection 5.4C on realizations with A in diagonal form). (a) Show that H(s) can be realized as the sum or the parallel combination of realizations of terms in the form —^-—; i.e., H(s) is realized by the system represented in the block diagram of Fig. 5.4, where r/ = QZ?/, / = 1 , . . . , n, with cj G /?", bi E /?", , and q = didt denotes the integrator circuit shown in Fig. 5.5. Note that
with q-
Xi
this realization of H{s) has advantages with respect to sensitivity to parameter vari ations.
1
"'
<=1
(7-X1
1
^
^n
q-K
FIGURE 5.4 Block diagram of the system in Example 5.5a
cVr
")""
)
I
^T
1
^i
FIGURE 5.5 Block diagram of an integrator
(b) Show that when poles are repeated, as, e.g., in the transfer function
427
H(s) =
-"TT + + ;r, then a parallel realization is as shown in the (5+1)2 ^+1 5+ 2 ^ block diagram given in Fig. 5.6.
1 qr+2 W
1
u
1
q+^
(7+1
\
J
FIGURE 5.6 Block diagram of the system in Example 5.5b (c) When there are complex conjugate poles, as, e.g., in the transfer function H(s) as + b al{s + c) (b - ac)/(s + cf J, then H(s) can be written as H(s) (S + C)2 + ^2^—-^-^-^ --^-^ I + (j2/(^ + c)2) •' 1 + (dVis + C)2) and can be realized as indicated in the block diagram given in Fig. 5.7.
" ^
i
k
1 q +c
1 g +c
^
a
1
b- ac 1
c^•
k/
d^
FIGURE 5.7 Block diagrams of the system in Exercise 5.5c Note that in addition to sum or parallel realizations discussed in (a) to (c), there 5+1 are also product or cascade realizations where, for example, H(s) (5 + 2)(5 + 3) 5+ 1 1 1 5+ 1 is realized as 5 + 2 5 + 3' 5+ 2 5+ 3 5.6. Consider the transfer function H(s) =
5+ 3 5+ 1 5
•5 + 3 5 + 1 (a) Determine the pole polynomial of H(s) and the McMillan degree of H(s). (b) Determine a minimal realization {A, B, C, D] of H(s), where A is a diagonal matrix. 5.7. Consider a system described by equations of the form y = [Ci, Ci]
0 [0
Xi
AIJL^IJ
Bi\
, where Ai is the Jordan block associated with the eigenvalue A, and Ai
CHAPTER 5 :
Reahzation Theory and Algorithms
428
is the complex conjugate of Ai associated with A. Show that the similarity transformation
Linear Systems
matrix P given by P = .
. can be used to reduce the representation given above
to one that involves only real system parameters, namely, Xi X2\
ReAi -ImAi
Im A\ ReAi
y = [ReCi,
2 Re Bi -llmBi
ImCi\
Note that this is a way of obtaining representations involving only real coefficients from realizations with A in diagonal or in Jordan form that may contain complex numbers. 5.8. Determine an observable reaHzation oiH{s) given in Exercise 5.1, where the matrix A is in block companion form, by using (a) the numerator polynomial matrix, (b) the Hankel matrix. s+2 5.9. Show that if the system {A, B, C, D} reaHzes H{s) = -^—;r and \sl - A\ = s^ + s^ + 2s -\- \ 2 ^ + 1 , then (A, B) must be controllable and (A, C) must be observable. Hint: Use Theorem 3.11. 5.10. Show that if the system {A, B, C, D] realizes the transfer function matrix H{s) and 1^/ - A| = mnis), the monic least common denominator of the entries of H(s), then {A, B, C, D} is a minimal realization of H(s). Can one always find a realization of ^(^) that satisfies this property? Hint: Use Theorem 3.11. 5.11. Consider a scalar proper rational transfer function H(s) = n(s)/d(s), and let x = AcXc + BcU, y = CcXc + DcU be a realization of H(s) in controller form (see Subsection 5.4B). (a) Show that the realization {A^, B^ Cc, Dc} is always controllable. (b) Show that {Ac, Be, Cc, Dc] is observable if and only if n{s) and d{s) do not have any factors in common, i.e., they are prime polynomials. (c) State the dual results to (a) and (b) involving a realization in observer form. 5.12. Consider 3. p X m proper rational transfer function matrix H(s) and the algorithm that leads to a reaHzation {Ac, Be, Cc, Dc} of H(s) = N(s)D(s)~^ in controller form [see (4.22) and (4.24)]. (a) Show that (Ac, Cc) is observable if and only if rank '
' = m for any A complex N(\)\ scalar. Hint: Use the eigenvalue/eigenvector tests for observability. Note that this rank condition is a necessary and sufficient condition for N(s) and D(s) to be right coprime (see Theorem 2.4 in Subsection 7.2D). (b) State the dual result to (a) that involves a realization in observer form {Ao, Bo, Co, Do) of His) = D-\s)N(s). ^ -i^
X
rr^ X
^(S)
S'^ -
S +
1
. Determine a realization in controller 5.13. Let H(s) = ^ K = -^ z T + s- 1 ^5 -minimal? ^4 + s^ Explain your answer. Hint: Use the results of Exerform. Is your d(s) realization cise 5.11.
5.14. For the transfer function H(s) = (a) (b) (c) (d)
429
s+ 1 ,find ^2 + 2 '
CHAPTERS:
an uncontrollable realization, an unobservable realization, an uncontrollable and unobservable realization, a minimal realization.
Realization Theory and Algorithms
5.15. Find minimal discrete-time state-space realizations of the transfer function matrix -z + 2 1 z+ 3 z+1 H(z) = z z+1 -z+1 z + 2using all the realization methods described in this chapter. 5.16. Given is the system depicted in the block diagram of Fig. 5.8, where H(s) = s^ + 1 — — —. Determine a minimal state-space representation for the closed{s + 1)(^ + 2){s + 3) F F loop system, using two approaches. In particular:
^ : ^
u
His)
L
y
FIGURE 5.8 Block diagram of the system in Exercise 5.16
(a) First, determine a state-space realization for H(s), and then, determine a minimal state-space representation for the closed-loop system. (b) First, find the closed-loop transfer function, and then, determine a minimal statespace representation for the closed-loop system. Compare the two approaches. 5.17. Consider the system depicted in the block diagram of Fig. 5.9, where H(s) = s+ I k — — and G(s) = with k,a EL R. Presently, H{s) could be viewed as the ^(^ + 3) s+a system to be controlled and G{s) could be regarded as a feedback controller.
r^ V ".
H{s)
-J
t
G(s)
y
FIGURE 5.9 Block diagram of the system in Exercise 5.17
(a) Obtain a state-space representation of the closed-loop system by (i) first, determining realizations for H{s) and G{s) and then combining them; (ii) first, determining Hds), the closed-loop transfer function. (b) Are there any choices for the parameters k and a for which your closed-loop statespace representation is uncontrollable and unobservable? If your answer is yes, specify. 5.18. For H{s) as in Exercises 5.1, 5.2, and 5.3, determine internally balanced minimal realizations, using singular-value decomposition.
430
5.19.
Let 1 ^2 + 1
Linear Systems H{s)
5^ + 1
1 5+ 2
5^ + 1
7T3 5^ + 1 ^3 + 2 5 + 1
^3 + 1 be a transfer function matrix. (a) Find a minimal realization for H(s) in controller form. (b) Find an internally balanced minimal realization for H{s). ^4
5.20.
Consider the controllable (-from-the-origin) and observable system given by i = Ax-\-Bu, y = Cx-\- Du and its equivalent representation x = Ax-\-Bu, y = Cx-\- Du, where A = PAP-\ B = PB,C = CP~\ and D = D. Let Wr and Wo denote the reachability and observability Gramians, respectively. (a) Show that Wr = PWrP* and Wo = {p-^yWoP~\ where P* denotes the complex conjugate transpose of P. Note that P* = P^ when only real coefficients in the system equations are involved. Using singular-value decomposition (refer to the Appendix), write Wr = Ur^rV; where WU =I,VV values of W. Define
and
Wo = Uo^oV;,
/, and E = diag (Ei,E2,... ,S„) with E/ the singular
H=iY}J'yu:Ur{Y}J').. and using singular-value decomposition, write
where U^UH '-1J VHVH = ^' Prove the following: ,l/2x-l (b) IfP = Pin = y / / ( E ; / ^ ) " > ; , t h e n W ; = / , W ^ , : (c) If P = Pout = Ufj(E^)
V,*, then Wr = ^]i,Wo-= 1.
(d) If P = Pib = Pin^ll^ = ^]i^Pouu then Wr = Wo= S//. Note that the equivalent representations {A,B,C,D} in (b), (c), and (d) are called, respectively, inputnormal, output-normal, and internally balanced representations. 5.21.
Show that the representation (4.61) in Section 5.4 is internally balanced, in view of Exercise 5.20(d).
5.22.
Consider the transfer function H{s)
4^3 + 3 5 + 1
and determine a minimal
realization for H{s). 5.23.
Transform the system " -1 0 1
2 1 0
-2" 1 -1
x+
"2" 0 u, 1
};=[l,l,0]x
into Xc = AcXc -\-BcU, y = CcXc, an equivalent representation in controller form utilizing the following two methods: (a) Use of a similarity transformation. (b) Determination of the transfer function H{s) and use of a realization algorithm.
5.24. Consider the two system 5*1 and 5*2 in series as depicted in the block diagram given in Fig. 5.10. Suppose that 5*1 is described by yi{s) = Hi{s)Ui{s), where ^i('^) ^
^ + 1' ^^^ similarly, that 5*2 is described by 3^2('^) = ti2(s)u2(s), where 5+ 2 U2{s)=yi{s) and7/2(5) = ——-.
" ~i^
1
^1
y,
1 V
=u^
^1
>2 ^2
*
1 l_
FIGURE 5.10 Block diagram of the system in Exercise 5.24
(a) Determine controllable and observable state-space descriptions for the individual subsystems 5*1 and 5*2. (b) Determine a state-space representation for the entire system S, using your answer in (a). (c) Is the state-space description of S in (b) controllable? Is it observable? What is the transfer function H{s) of 5*? 5.25. Consider a system described by 2
1 (5+1)2
0
5+1
^1(5) U2{s)
(a) What is the order of a controllable and observable realization of this system? (b) If we consider such a realization, is the resulting system controllable from the input W2? Is it observable from the output yi ? Explain your answers. 5.26. Consider the system described by H{s) 5-1
5+ 2
1 5 - ( l + e)
(y(s) =H(s)u(s))
and C(5)
(^(5) = C{s)f{s)) connected in series (e G /?).
(a) Derive minimal state-space realizations for H{s) and C{s) and determine a (second-order) state-space description for the system y{s) = H{s)C{s)f{s). (b) Let e = 0 and discuss the implications regarding the overall transfer function and your state-space representations in (a). Is the overall system now controllable, observable, asymptotically stable? Are the poles of the overall transfer function stable? [That is, is the overall system BIBO stable? (See Chapter 6.)] Plot the states and the output for some nonzero initial condition and a unit step input and comment on your results. (c) In practice, if H{s) is a given system to be controlled and C{s) is a controller, it is unlikely that e will be exactly equal to zero and therefore the situation in (a), rather than (b), will arise. In view of this, comment on whether open-loop stabilization can be used in practice. Carefully explain your reasoning.
431 CHAPTER 5 :
Realization Theory and Algorithms
CHAPTER 6
Stability
Dynamical systems, either occurring in nature or manufactured, usually function in some specified mode. The most common such modes are operating points that frequently turn out to be equilibria. In this chapter we will concern ourselves primarily with the qualitative behavior of equilibria. Most of the time, we will be interested in the asymptotic stability of an equilibrium (operating point), which means that when the state of a given system is displaced (disturbed) from its desired operating value (equilibrium), the expectation is that the state will return to the equilibrium. For example, in the case of an automobile under cruise control, traveling at the desired constant speed of 50 mph (which determines the operating point, or equilibrium condition), perturbations due to hill climbing (hill descending), will result in decreasing (increasing) speeds. In a properly designed cruise control system, it is expected that the car will return to its desired operating speed of 50 mph. Another qualitative characterization of dynamical systems is the expectation that bounded system inputs will result in bounded system outputs, and that small changes in inputs will result in small changes in outputs. System properties of this type are referred to as input-output stability. Such properties are important, for example, in tracking systems, where the output of the system is expected to follow a desired input. Frequently, it is possible to establish a connection between the input-output stability properties and the Lyapunov stability properties of an equilibrium. In the case of linear systems, this connection is well understood. 6.1 INTRODUCTION In this chapter we present a brief introduction to stability theory. We are concerned primarily with linear systems and systems that are a consequence of linearizations 432
of nonlinear systems. As in the other chapters of this book, we consider finitedimensional continuous-time systems and finite-dimensional discrete-time systems described by systems of first-order ordinary differential equations and systems of first-order ordinary difference equations, respectively. We will consider both internal descriptions and external descriptions of systems. For the former, we present results for various types of Lyapunov stability of an equilibrium (under assumptions of no external inputs), while for the latter we develop results for input-output stability (under assumptions of zero initial conditions). We also present results that connect these two stability types. By considering both Lyapunov stability and input-output stability, we are frequently able to conduct a more complete qualitative analysis of a system than can be accomplished by applying only one type of stability analysis. Recall that in Subsections 2.4C and 2.7E we have already encountered stability properties of an equilibrium for time-invariant, continuous-time, and discrete-time systems, respectively.
A. Chapter Description This chapter is organized into three parts. In Part 1 (Sections 6.3 through 6.8) we address the Lyapunov stability of an equilibrium, in Part 2 (Section 6.9) we consider input-output stability, and in Part 3 (Section 6.10), we treat both Lyapunov stability and input-output stability of (time-invariant) discrete-time systems. Part 1 is preceded by Section 6.2, which contains some background material from linear algebra. In Section 6.2 we provide additional background in linear algebra dealing with bilinear functional and congruence, Euclidean vector spaces, and linear transformations on Euclidean vector spaces. This material constitutes a continuation of the material presented in Subsections I.IOA and I.IOB and Section 2.2, and the notation used in those sections will also be employed in the second section of this chapter. In Section 6.3 we introduce the concept of equilibrium of dynamical systems described by systems of first-order ordinary differential equations, and in Section 6.4 we give definitions of various types of stability in the sense of Lyapunov (including stability, uniform stability, asymptotic stability, uniform asymptotic stability, exponential stability, and instability). In Section 6.5 we establish conditions for the various Lyapunov stability and instability types enumerated in Section 6.4 for linear systems {LH), (L), and (P), Most of these results are phrased in terms of the properties of the state transition matrix for such systems. In Section 6.6 we state and prove necessary and sufficient conditions for the exponential stability of the equilibrium of nth-order ordinary differential equations with constant coefficients and systems of linear first-order ordinary differential equations (L). These involve geometric criteria (the interlacing theorem) and algebraic criteria (the Routh-Hurwitz criterion). In Section 6.7 we introduce the Second Method of Lyapunov, also called the Direct Method of Lyapunov, to establish necessary and sufficient conditions for various Lyapunov stability types of an equilibrium for linear systems (L). These results, which are phrased in terms of the system parameters [coefficients of the matrix A for system (L)], give rise to the Lyapunov matrix equation.
433 CHAPTER 6
Stability
434 Linear Systems
In Section 6.8 we use the Direct Method of Lyapunov in deducing the asymptotic stabihty and instabiUty of an equihbrium of nonhnear autonomous systems (A) from the stabihty properties of their hnearizations. In Section 6.9 we estabhsh necessary and sufficient conditions for the inputoutput stabihty (more precisely, for the uniform bounded input/bounded output stability) of continuous-time, linear, time-varying systems, and linear, time-invariant systems. These results involve the system impulse response matrix. In this section we also establish a connection between the bounded input/bounded output stability of linear systems and the exponential stability of an equilibrium of linear systems. The stability results presented in Sections 6.3 through and including Section 6.9 pertain to continuous-time systems. In Section 6.10 we present analogous stability results for discrete time systems; however, in the interests of economy, we confine ourselves in this section to time-invariant systems. Also, to give the reader a glimpse into the qualitative theory of dynamical systems described by nonlinear equations, we present in this section Lyapunov stability results for finite-dimensional dynamical systems described by systems of nonlinear first-order ordinary difference equations. We conclude the chapter with comments concerning some of the existing literature dealing with the present topic. As in all the other chapters, problems are provided at the end of the chapter to further clarify the subject at hand.
B. Guidelines for the Reader In a first reading, the background material on linear algebra, given in Section 6.2, can be reviewed rather quickly, as needed. In a first course on linear systems, the reader needs to acquire familiarity with the notion of an equilibrium (Section 6.3) and various stability concepts of an equilibrium (Section 6.4). Such a course may be confined to studying the stability properties of an equilibrium for time-invariant systems (L) (refer to Theorem 5.6 in Section 6.5). This may be followed by coverage of Section 6.7, where the principal Lyapunov stability results for time invariant systems (L) are established in terms of the properties of the Matrix Lyapunov Equation. In a first reading. Section 6.8 should also be covered in its entirety, where conditions are established that enable one to deduce the stability properties of a time-invariant nonlinear system (A) from the hnearization of (A). The reader should concentrate on the input-output stability results for linear, time-invariant, continuous-time systems given in Theorems 9.4 and 9.5 in a first course on linear systems. Finally, in a first reading, the reader may consider some or all of the counterparts of the above results for the case of linear, time-invariant, discrete-time systems developed in Section 6.10. 6.2 MATHEMATICAL BACKGROUND MATERIAL In this section we provide additional background material from linear algebra and matrix theory. As in previous sections where we present such material, we assume
that the reader has some background in these areas, and therefore, our presentation will be in the form of a summary, rather than a development of the subject at hand. This section consists of three subsections. In the first subsection we address bilinear functionals and congruence, in the second subsection we present material on Euclidean vector spaces, and in the third subsection we address some issues dealing with linear transformations on Euclidean spaces.
A. Bilinear Functionals and Congruence We consider here the representation of bilinear functionals on real finite-dimensional vector spaces. Throughout this subsection, V is assumed to be an /i-dimensional vector space over the field of real numbers R. We define a bilinear functional on V as a mapping f : V X V -^ R having the properties f(av^ + jSv^ w) = af(v\
w) + )8/(v^ w)
(2.1)
/(v, yw^ + 8w^) = yf(v, w^) + 8f(v, w^)
(2.2)
for all a, I3,y,d G R and for all v^ v^, w^, w^ in Y. A direct consequence of this definition is the more general property
for all aj, ^^ G R and v^, w^ ^ V, j = 1 , . . . , r,-and k = 1 , . . . , ^. Now let {v^ . . . , v"} be a basis for the vector space V and let fij = f(y\vJ),
i,j = l , . . . , n .
(2.3)
The matrix F = [ftj^ is called the matrix of the bilinear functional f with respect to The characterization of bihnear functionals on real finite-dimensional vector spaces is given in the following result, which is a direct consequence of the above definitions and the properties of bases: l e t / be a bilinear functional on V and let {v^ . . . , v"} be a basis for V. Let F be the matrix of the bilinear functional/ with respect to the basis {v^ . . . , v'^}. If x and y are arbitrary vectors in V, and if ^ and 7] are their coordinate representations with respect to the basis {v^ . . . , v"}, then
f{x, y) = eFv = XJlM^^J'
(2.4)
/-I 7 = 1
Conversely, if we are given any nX n matrix F, we can use (2.4) to define the bilinear functional/ whose matrix with respect to the given basis {v^ . . . , v"} is, in turn, F again. In general it therefore follows that on finite-dimensional vector spaces, bilinear functionals correspond in a one-to-one fashion to matrices. The particular one-to-one correspondence depends on the particular basis chosen. A bilinear functional/ on V is said to be symmetric if f(x, y) = f(y, x) for all x,yGV and skew symmetric if f(x, y) = — f{y, x) for all x, y E V.
435 CHAPTER 6:
Stability
436 Linear Systems
For symmetric and skew symmetric bilinear functional the following results are easily proved: let {v^ . . . , v^} be a basis for V, and let F be the matrix for a bilinear functional with respect to {v^ . . . , v'^}. Then 1. / is symmetric if and only if F = F^. 2. f is skew symmetric if and only if F = -F^. 3. For every bilinear functional/, there exists a unique symmetric bilinear functional /i and a unique skew symmetric bilinear functional fz such that / = /i + fi.
(2.5)
We call /i the symmetric part off and /2 the skew symmetric part of/. The above result motivates the following definitions: an n X n matrix F is said to ht symmetricif F = F^ dind skew symmetric if F = —F^. Using definitions, the following result is easily established: l e t / be a bilinear functional on V and let / and /2 be the symmetric and skew symmetric parts of/, respectively. Then
and
/i(v,w) = ^[/(v,w) + /(w,v)]
(2.6)
/2(v, w) = \ [/(v, w) - f{w, v)]
(2.7)
for all V, w G V. Next, we define the quadratic form induced by a bilinear functional/ on V as f(y) = /(^^ v) for all V E y. It is easily verified that in terms of matrices we have
fix) = fF^ = ^JlfiMp
(2-8)
where ^ denotes the coordinate representation of x with respect to the basis {v^...,v"}. For quadratic forms we have the following result: l e t / and g be bilinear functional on V. The quadratic forms induced b y / and g are equal if and only iff and g have the same symmetric part. In other words, /(v) = g(v) for all v E V if and only if k [/(v, w) + /(w, V)] = i [g(v, w) + g(w^ V)]
(2.9)
for all v,w EiV. From this result we can conclude that when treating quadratic functional, it suffices to work with symmetric bilinear functionals. It is also easily verified from definitions that a bilinear functional/ on a vector space V is skew symmetric if and only if /(v, v) = 0 for all v G V. Next, l e t / be a bilinear functional on a vector space V, let {v\ . . . , v'^} be a basis for y, and let F b e the matrix off with respect to this basis. Let {v^ . . . , v'^} be another basis whose matrix with respect to {v^ . . . , v'^} is P. It can readily be verified that the matrix F off with respect to the basis {v^ . . . , v"} is given by F = P^FP.
(2.10)
This result gives rise to the following concept: annX n matrix F is said to be congruent to an nXn matrix F if there exists a nonsingular matrix P such that (2.10) holds. We express congruence of two matrices F, F by writing F--F.lt is easily shown that ~ is reflexive, symmetric, and transitive, and as such it is an equivalence relation.
For practical reasons we are interested in determining the simplest matrix congruent to a given matrix, or what amounts to the same thing, the most convenient basis to use in expressing a given bilinear functional. If, in particular, we confine our interests to quadratic functional, then it suffices, as observed earlier, to consider symmetric bilinear functionals. The following result, called Sylvester's Theorem, addresses this issue. The proof of this result is rather lengthy, and the interested reader should consult one of the references on linear algebra cited at the end of this chapter for details. L e t / be any symmetric bilinear functional on a real n-dimensional vector space V. Then there exists a basis {v^,...,v^}ofV such that the matrix of/ with respect to this basis is of the form 1
\n.
(2.11)
-1
0
The integers r and p in this matrix are uniquely determined by the bilinear form. Sylvester's Theorem allows the following classification of symmetric bilinear functionals: the integer r in (2.11) is called the rank of the symmetric bilinear functional/, the integer p is called the index of/, and n is called the order off. The integer s = 2p-r (i.e., the number of + I's minus the number of - 1 's) is called the signature of/. A bilinear functional / on a vector space V is said to be positive (or positive semidefinite) if /(v, v) > 0 for all v E V and strictly positive (or positive definite) if /(v, v) > 0 for all V 7^ 0, V G y [note that /(v, v) = 0 for v = 0]. It is readily verified that a symmetric bilinear functional is strictly positive if and only \i p = r = n in (2.11) and positive if and only if p = r. Finally, we say that a bilinear functional/ is negative (or negative semidefinite) if - / is positive and strictly negative (or negative definite) if - / is strictly positive. Also, a bilinear functional/ is said to be indefinite if in (2.11) r > /? > 0.
B. Euclidean Vector Spaces As in the preceding subsection, we assume throughout this subsection that Visa real vector space. A bilinear functional / defined on V is said to be an inner product if (i) / is symmetric and (ii) / is strictly positive. A real vector space V on which an inner product is defined is called a real inner product space, and a real finite-dimensional vector space on which an inner product is defined is called a Euclidean space. Since we will always be concerned with a given bilinear functional on V, we will write (v, w) in place of f(v, w) to denote the inner product of v and w. Accordingly,
437 CHAPTER 6:
Stability
438
the axioms of a real inner product are given as
Linear Systems
^ 2. 3. 4.
(v, v) > 0 for all v 7^ 0 and (v, v) = 0 when v = 0. (v, w) = (w, v) for all v, w e V. (av + ^w, u) = a(v, u) + ^{w, u) for all u,v,w ^V and all a, (3 ^ R. (w, av + j8w) = a(M, v) + J8(M, W) for all u,v,w G V and all a, (3 G R.
In the following we enumerate several results on Euclidean spaces, all of which are easily proved: 1. 2. 3. 4.
The inner product (v, w) = 0 for all v E F if and only if w = 0. Let ^ G L(V, V), Then (v, ^ w ) = 0 for all v^wGVif and only if ^ = 0. Let ^ , gS G L(K V). If (v, siw) = (v, ^w) for all v, w G V, then ^ = ^. Let A G /^"x^ If ^^AT; = 0 for all ^, r^ G 7?^ then A = 0. We now define the function ||v|| = (v,v)^/2
(2.12)
for all V G y. It is easily verified (using definitions and the properties of bilinear functionals) that this function satisfies the axioms of a norm (refer to Subsection I.IOB), i.e., for all v, w in V and for all scalars a, the following hold: L ||v|| > 0 for all V 7^ 0 and ||v|| = 0 when v = 0. 2. ||a:v|| = |«| • ||v||, where \a\ denotes the absolute value of the scalar a. 3. ||v + w\\ < ||v|| + ||w||. In the usual proof of 3, use is made of the Schwarz Inequality, which states that \{y,w)\ < IMI'IMI
(2.13)
for all V, w G V, and |(v, w)| = ||v|| • IHI if and only if v and w are linearly dependent. Another result for Euclidean spaces that is easily proved is the parallelogram law, which asserts that for all v,w G V, the equality ||v + wf + ||v - w|p = 2||v|p + 2|Hp.
(2.14)
Before proceeding, we recall that a vector space V on which a norm is defined is called a normed linear space. It is therefore clear that Euclidean vector spaces are normed linear spaces with norm defined by (2.12) (refer to Subsection l.lOB). We also recall that any norm can be used to define a distance function, called a metric, on a normed vector space by letting d{u,v) = \\u-v\\
(2.15)
for all w, V G V (refer to Subsection l.lOB). Using thepropertiesofnorm, it is readily verified that for all w, v, w G V, it is true that 1. d(u, v) = d(v, u). 2. d{u, v) > 0 and d{u, v) = 0 if and only if w == v. 3. d{u, v) < d{u, w) + d{w, v).
Without making use of inner products or norms, one can define a distance func- 439 tion having the properties enumerated above on an arbitrary set. Such a set, along CHAPTER 6: with the distance function, is then called a metric space. It is thus clear that Euclidean Stabihty vector spaces are metric spaces as well. The concept of inner product enables us to introduce the notion of orthogonality: two vectors v,w ^V are said to be orthogonal (to one another) if (v, w) = 0. This is usually written as v ± w. With the aid of the above concept we can immediately establish the famous Pythagorean Theorem: for v, w G V, if v _L w, then ||v + w|p = ||v|p + ||w|p. A vector v E V is said to be a unit vector if ||v|| = 1. Let w T^ 0, and let u = (l/||w||)w. Then the norm of u is ||w|| = (l/||w||)||w|| ^ 1, i.e., w is a unit vector. We call the process of generating a unit vector from an arbitrary nonzero vector w normalizing the vector w. Now let {w^,..., w^} be an arbitrary basis for V and let F = [fij] denote the matrix of the inner product with respect to this basis, i.e., fij = (w^ w^) for all / and j . More specifically, F denotes the matrix of the bilinear functional/ that is used in determining the inner product on V with respect to the indicated basis. Let ^ and rj denote the coordinate representation of vectors x and y, respectively, with respect to {w^ . . . , w""}. Then we have by (2.4), (X, y) = ^Fv = V^F^ = XXfiJ^J^i'
(2.16)
Now by Sylvester's Theorem [see (2.11) and the discussion following that theorem], since the inner product is symmetric and strictly positive, there exists a basis {v^ .. .yV^} for V such that the matrix of the inner product with respect to this basis is the nX n identity matrix /, i.e., ' 0, if / 7^ J, (v\vJ) = Sij = 1 . . . . (2.17) 1, lii = J. This motivates the following definition: if {v\ . . . , v"} is a basis for V such that (v^, v^) = 0 for all / T^ J , i.e., if v^ J_ v^ for all / T^ j , then {v^ . . . , v'^} is called an orthogonal basis. If in addition, (v^ v^ = 1, i.e., ||v^|| = 1 for all /, then {v^ . . . , v"} is said to be an orthonormal basis for V [thus, {v^ . . . , v"} is orthonormal if and only if{v\vJ)^8ijl Using the properties of inner product and the definitions of orthogonal and orthonormal bases, we can easily establish several useful results: 1. Let {v^ . . . , v^} be an orthonormal basis for V. Let x and y be arbitrary vectors in y, and let the coordinate representation of x and y with respect to this basis be ^^ = (^1, ...,^n) and 7]^ = (171,..., 7]n), respectively. Then (x,y) = ev and
= V^^ =X^iVi
ll^ll = (e^f^
= 7^f + - - - + e
(2.18) (2.19)
2. Let {v^ . . . , v^} be an orthonormal basis for V and let x be an arbitrary vector. The coordinates of x with respect to {v^ . . . , v"} are given by the formula ^. = (x,v'),
i = l...,n.
(2.20)
440 Linear Systems
3. Let {v^ . . . , v"} be an orthogonal basis for V. Then for all x E V, we have /^ ^l\ /^ ^n\ . = ( ^ v ' + - + i ^ v « . (2.21) 4. {ParsevaVs identity) Let {v^ . . . , v"} be an orthogonal basis for V. Then for any x,y ^V, we have
(„) = X ^ ^ .
(2.22)
5. Suppose that x^,..., x^ are mutually orthogonal nonzero vectors in V, i.e., x^ ± ;c^, / ^ 7. Then x^ . . . , x^ are linearly independent. 6. A set of k nonzero mutually orthogonal vectors is a basis for V if and only if k = dimV = n. 7. For V there exist not more than n mutually orthonormal vectors (called a complete orthonormal set of vectors). 8. {Gram-Schmidtprocess) Let {w^,..., w"} be an arbitrary basis for V. Set 1
1
1
^ ^ ^
M
^
n i l (2.23)
«-i V
==
ii—
Then {v^ . . . , v^} is an orthonormal basis for V. 9. If v^ . . . , v^, k < n, are mutually orthogonal nonzero vectors in V, then we can find a set of vectors v^"*"^ . . . , v^ such that the set {v\ . . . , v"} forms a basis forV. 10. {BesseVs inequality) If {w^ . . . , w^} is an arbitrary set of mutually orthonormal vectors in V, then k
^ |(w, wOP ^ IMP
(2.24)
/= i
for all w E.V. Moreover, the vector k
u = w — ^_^(w, w^)w^ is orthogonal to each w\i = 1 , . . . , k. 11. Let ]¥ be a linear subspace of V, and let W^ = {v^V
:(v,w) = 0 for all w G W}.
(2.25)
(i) Let {w^,..., w^} span W. Then v G W-^ if and only if v ± w^ for j = l,...,yt. (ii) W^ is a linear subspace of V. (iii) ^ = dimV = dimT^ 4- dimW^. (iv) (W^)-^ = W. (v) V = WeW^.
(vi) Let u,v ^ V. If u = u^ -\- u^ and v = v^ + v^, where u^, v^ G W and u^, v^ G W^, then {u,v) = (u\v^) and
+ (u^,v^)
||M|| = V||wi|P + \\uY.
(2.26) (2.27)
We conclude this subsection with the following definition: let W be a linear subspace of V. The subspace W^ defined in (2.25) is called the orthogonal complement ofW,
C. Linear Transformations on Euclidean Vector Spaces In this subsection we consider some of the properties of three important classes of linear transformations defined on Euclidean spaces: orthogonal transformations, adjoint transformations, and self-adjoint transformations. Unless otherwise explicitly stated, V will denote an n-dimensional Euclidean vector space. Let {v^...,v"} be an orthonormal basis for V, let v^ = ^1=\ PjiV^,i = 1 , . . . , /2, and let P denote the matrix determined by the real scalars pij. The following question arises: when is the set {v^ . . . , v"} also an orthonormal basis for VI To determine the desired properties of P, we consider (v^ vO = ( X Pki^'^ S PuA = X PkiPiM. v^ \k=i 1=1 J k,i
(2.28)
So that {v\ v^) = 0 for / T^ j and (v^ v^) = 1 for i = j , we require that n
{v\vJ) = X k,i = i
n
PkiPijSki = ^PkiPkj
= Sij,
(2.29)
k=i
i.e., we require that P^P = I, (2.30) where, as usual, / denotes the n X n identity matrix. The above discussion is summarized in the following result: let {v^ . . . , v"} be an orthonormal basis for V and let v^ = X}= i Pji^jy i = \,.. .,n. Then {v^ . . . , v"} is an orthonormal basis for V if and only if P^ = P~^. This result, in turn, gives rise to the concept of orthogonal matrix. Thus, a matrix P G R^^^ such that P^ = P~^, i.e., such that P^P = P~^P = /, is called an orthogonal matrix. It is not difficult to show that if P is an orthogonal matrix, then either det P = \ or det P = -I. Also, if P and Qsire nX n orthogonal matrices, then so is PQ. The nomenclature used in the next definition will become clear shortly. We say that a linear transformation si from V into V is an orthogonal linear transformation if (^v, siw) = (v, w) for all v,w G V. We now enumerate several properties of orthogonal transformations. The proofs of these statements are straightforward. 1. Let si G L(V, V). Then ^ is orthogonal if and only if \\siv\\ = \\v\\ for all v G V. [Note that if si is an orthogonal linear transformation, then v _L w for all v, w G V if and only if siv 1 siw since (v, w) = 0 if and only if (^v, siw) = 0.]
441 CHAPTER 6:
Stability
442 Linear Systems
2. Every orthogonal linear transformation of V into V is nonsingular. 3. Let {v^ . . . , v"} be an orthonormal basis for V. Let ^ G L(V, V), and let A be the matrix of ^ with respect to this basis. Then M^ is orthogonal if and only if A is orthogonal. 4. Let d e L(V, V). If si is orthogonal, then det d = ± 1 . 5. Let si,^ E. L(V, V). If M and 2^ are orthogonal linear transformations, then 64SS is also an orthogonal linear transformation. The next result enables us to introduce adjoint linear transformations in a natural manner. Let ^ G L{V, V) and define g : V X V -^ Rby g(v, w) = (v, ^w) for all v,w GV. Then g is a bilinear functional on V, Moreover, if {v^ . . . , v'^} is an orthonormal basis for V, then the matrix of g with respect to this basis, denoted by G, is the matrix of ^ with respect to {v^ . . . , v"}. Conversely, given an arbitrary bilinear functional g defined on V, there exists a unique linear transformation ^ G L(V, V) such that (v, ^w) = g(v, w) for all v,w E. V, It should be noted that the correspondence between bilinear functionals and linear transformations determined by the relation (v, ^w) = g(v, w) for all v, w G V does not depend on the particular basis chosen for V; however, it does depend on the way the inner product is chosen for V at the outset. Now let ^ G L(V, V), set g(v, w) = (v, ^w), and let h(v, w) = g(w, v) = (w, ^v) = (^v, w). By the result given above, there exists a unique linear transformation, denote it by ^*, such that h(v, w) = (v, Ww) for all v,w G V. We call the hnear transformation ^* G L(V, V) the adjoint of^. We have the following results: 1. For each ^ G L(V, V), there is a unique «* G L(V, V) such that (v,^*w) = (^v, w) for all V, w G y. 2. Let {v\ . . . , v^} be an orthonormal basis for V, and let G be the matrix of the Hnear transformation ^ G L(V, V) with respect to this basis. Let G* be the matrix of ^* with respect to {v^ . . . , v"}. Then G* = G^. The above results allow the following equivalent definition of the adjoint linear transformation: let ^ G L(V, V). The adjoint transformation, ^*, is defined by the formula (v,^*w) = (^v,w)
(2.31)
for all v,w EV. In the following, we enumerate some of the elementary properties of the adjoint of linear transformations. The proofs of these assertions follow readily from definitions. Let ^ , gS G L(V, V), let d*, 9^* denote their respective adjoints, and let a be a real scalar. Then
1. 2. 3. 4.
(^y = M. (^ + my - 64* + m\ (asiT = aM\ (^SS)* = a*^*.
5. 6. 7. 8.
^* = S>, where ^ denotes the identity transformation. ©* = 0, where 0 denotes the null transformation. ^ is nonsingular if and only if ^4* is nonsingular. I f ^ is nonsingular, then (6^*)-^ = (si-^.
The following results (the proofs of which are straightforward) characterize orthogonal transformations in terms of their adjoints. Let si E L(V, V). Then si is orthogonal if and only if 6^* = si~^. Furthermore, si is orthogonal if and only if si~^ is orthogonal, and si~^ is orthogonal if and only if si* is orthogonal. Using adjoints, we now introduce two additional important types of linear transformations. Let si E L(V, V). Then si is said to be self-adjoint if si* = si, and it is said to be skew-adjoint if si* = -^. Some of the properties of such transformations are as follows. 1. Let si E L(V, V), let {v^ . . . , v'^} be an orthonormal basis for V, and let A be the matrix of si with respect to this basis. The following are equivalent: (i) si is self-adjoint, (ii) ^ is symmetric, (iii) (^v, w) = (v, siw) for all v,w E. V. 2. Let si E L(V, V), let {v\ . . . , v'^} be an orthonormal basis for V, and let A be the matrix of si with respect to this basis. The following are equivalent: (i) ^ is skew-adjoint, (ii) si is skew-symmetric, (iii) (^v, w) = -(v, siw) for all v, w E V. The next result follows from part (iii) of the above result. Let ^ E L(V, V), let {v^ . . . , v'^} be an orthonormal basis for V, and let A be the matrix of si with respect to this basis. The following are equivalent: (i) si is skew-symmetric, (ii) (v, siv) = 0 for all v E V. (iii) ^ v 1 V for all v E V. The following result enables one to represent arbitrary linear transformations as the sum of self-adjoint and skew-adjoint transformations. Let si E L(V, V). Then there exist unique ^di, ^ 2 ^ L(V,V) such that ^ = sii + ^2, where ^i is selfadjoint and si2 is skew-adjoint. This has the direct consequence that every real nXn matrix can be written in one and only one way as the sum of a symmetric and skewsymmetric matrix. The next result is applicable to real as well as complex vector spaces. (We will state it for complex spaces.) Let y be a complex vector space. Then the eigenvalues of a real symmetric matrix A are all real. If all eigenvalues of A are positive (negative), then A is called po^itive (negative) definite. If all eigenvalues of A are nonnegative (nonpositive), then A is cailod positive (negative) semidefinite. If A has positive and negative eigenvalues, then A is said to be indefinite. Next, let A be the matrix of a linear transformation d^ E L(V,V) with respect to some basis. If A is symmetric, then as indicated above, all its eigenvalues are real. In this case d- is self-adjoint and all its eigenvalues are also real; in fact, the eigenvalues of si and A are identical. Thus, there exist unique real scalars Ki,..., \p, p ^ n, such that det (si - X3) = det (A - XI) = (Ai - A)^i(A2 " A)'^^... (A;, - A)'^^
(2.32)
We summarize the above observations in the following. Let si E L(V, V). If si is self-adjoint, then si has at least one eigenvalue. Furthermore, all eigenvalues of si
443 CHAPTER 6:
Stability
444 Linear Systems
are real and there exist unique real numbers A i , . . . , A^, /? < n, such that Eq. (2.32) holds. As in Eq. (3.17) of Chapter 2, we say that in (2.32) the eigenvalues A/, / = 1 , . . . , p < n, have algebraic multiplicities mi, i = 1 , . . . , /?, respectively. Next, we examine some of the properties of the eigenvalues and eigenvectors of self-adjoint linear transformations. The proofs of these assertions are straightforward and follow mostly from definitions. Let si G L(V, V) be a self-adjoint transformation, and let Xi,..., Xp, p ^ n, denote the distinct eigenvalues of ^4. If v^ is an eigenvector for A/ and if v^ is an eigenvector for \j, then v' J_ v^ for all i y^ j . Now let ^ G L(V, V) and let A/ be an eigenvalue of ^ . Recall that Xi denotes the null space of the linear transformation d- - A/^, i.e.. Xi = {vGV
:(d-
A/J^)v = 0}.
(2.33)
Recall also that Mi is a linear subspace of V. From the last result given above, the following result follows immediately. Let d G L(V, V) be a self-adjoint transformation and let A/ and A^ be eigenvalues of 5i. If A/ 7^ Ay, then Mi 1 Mj. The proof of the next result is somewhat lengthy and involved. The reader should consult the references on linear algebra cited at the end of this chapter for details. Let si G L(V, V) be a self-adjoint transformation, and let A i , . . . , A^, p < n, denote the distinct eigenvalues of si. Then dim V = n = dim^Ti + dim>r2 + . . . + dim^Tp.
(2.34)
The next two results are direct consequences of the above result. 1. Let ^ G L(V, V). If d is self-adjoint, then (i) there exists an orthonormal basis in V such that the matrix of si with respect to this basis is diagonal; (ii) for each eigenvalue A/ of si we have dim>f/ = algebraic multiplicity of A/. We note that in the above theorem, the matrix A of the linear transformation si with respect to the chosen orthonormal basis in V is given by
A2
A =
A2
(2.35)
2. Let A be a real nX n symmetric matrix. Then there exists an orthogonal matrix P such that the matrix A defined by
A - P-^AP = P^AP
(2.36)
is diagonal.
445 CHAPTER 6:
Stability For symmetric bilinear functionals defined on Euclidean vector spaces, we have the following result. Let f(v, w) be a symmetric bilinear functional on V. Then there exists an orthonormal basis for V such that the matrix of/ with respect to this basis is diagonal. For quadratic forms, the following useful result is easily proved. Let f(x) be a quadratic form defined on V. Then there exists an orthonormal basis for V such that if ^^ = (^i,..., ^„) is the coordinate representation of x with respect to this basis, then f(x) = ai^f-\ h a„^^ for some real scalars o^i,..., a„. The final result of this section, which we state next, is called the Spectral Theorem for self-adjoint linear transformations. For the proof of this result, the interested reader should consult one of the references on linear algebra cited at the end of this chapter. Before stating this theorem, we recall that a transformation 2P G L(y, V) is a projection on a linear subspace of V if and only if 2^^ = 2?* (refer to Subsection 2.2K). Also, for any projection SP, T = 91(2^) 0 >r(S?>), where 2/l(2P) is the range of 2^ and M{^) is the null space of 2^ (refer to Subsection 2.2K). Furthermore, we call 2^ an orthogonal projection if 2/1(9^) 1 >r(2^). The Spectral Theorem for self-adjoint linear transformations: let ^ E L{V, V) be a self-adjoint transformation, let A i , . . . , A^ denote the distinct eigenvalues of si, and let Mi be the null space oi d^ - Xi3. For each / = 1 , . . . , p, let ^t denote the projection on Xi along JV"-^. Then 1. ^i is an orthogonal projection for each / = 1 , . . . , /?. 2. ?Pi?Pj = 0 for / 7^ J, i,j = 1 , . . . , p, 3. Xy^^i Sy" = 3, where 3" ^ L(V,V) denotes the identity transformation.
4. ^ = S ; . i A # ; .
PARTI LYAPUNOV STABILITY 6.3 THE CONCEPT OF AN EQUILIBRIUM In this section we concern ourselves with systems of first-order ordinary differential equations (£"), i.e., X = fit, x\
(E)
where x G R^. When discussing global results, we shall assume that f : R^ XR^ ^ R^, while when considering local results, we may assume that f : R^ X B(h) -> R^ for some h> O.On some occasions, we may assume that t G R, rather than t ^ R^. Unless otherwise stated, we shall assume that for every (to, XQ), to G R'^, the initialvalue problem X = fit, x),
xito) = xo
il)
446 Linear Systems
possesses a unique solution (pit, to, XQ) that exists for all t > t^ and that depends continuously on the initial data (^o. -^o)- Refer to Chapter 1 for conditions that ensure that (/) has these properties. DEFINITION 3.1. A point Xe E R^ is called an equilibrium point of (E), or simply an equilibrium of (E) (at time f G R^), if fit, Xe) = 0 for all t> t.
m
We note that if Xe is an equilibrium of (£") at f, then it is also an equilibrium at all T ^ f. We also note that in the case of autonomous systems X = fix),
(A)
and in the case of T-periodic systems X = fit, X),
fit, X) = fit + T, X),
iP)
a point Xe E R^ is an equilibrium at some time f if and only if it is an equilibrium at all times. [Refer to Chapter 1 for the definitions of symbols in (A) and (P).] We further note that if Xe is an equilibrium at f of (£"), then the transformation s = t - t yields dx
r.
~ ^
and Xe is an equilibrium at 5" = 0 of this system. Accordingly, we will henceforth assume that f = 0 in Definition 3.1 and we will not mention f again. Furthermore, we note that for any to > 0, (f)it, to, Xe) = Xe
for all t > to,
i.e., the equilibrium Xe is a unique solution of iE) with initial data given by (/)(^, to, Xe) = Xe.
We will call an equilibrium point Xe of iE) an isolated equilibrium point if there is an r > 0 such that Bixe, r) C R^ contains no equilibrium point of iE) other than Xe itself. [Recall that Bixe, r) = {x E. R^ : \\x - Xe\\ < r}, where || • || denotes some norm defined on R^.] Unless stated otherwise, we will assume throughout this chapter that a given equilibrium point is an isolated equilibrium. Also, we will usually assume that in a given discussion, unless otherwise stated, the equilibrium of interest is located at the origin of R^. This assumption can be made without loss of generality by noting that if Xg T^ 0 is an equilibrium point of (£*), i.e., fit, Xe) = 0 for all t ^ 0, then by letting w = x - Xe,we obtain the transformed system w = Fit,w)
(3.1)
with Fit, 0) = 0 for all t > 0, where Fit, w) = fit, w + Xe).
(3.2)
Since (3.2) establishes a one-to-one correspondence between the solutions of iE) and (3.1), we may assume henceforth that the equilibrium of interest for iE) is located at the origin. This equilibrium, x = 0, will be referred to as the trivial solution of iE).
Before concluding this section, it may be fruitful to consider some specific cases.
447
EXAMPLES.1. In Example 4.4 in Chapter 1 we considered the simple pendulum given in Fig. 1.7. Letting xi = x and X2 = i in Eq. (4.12) of Chapter 1, we obtain the system of equations
CHAPTER 6:
Xi = X2
X2 = -/csinxi,
(3.3)
where ^ > 0 is a constant. Physically, the pendulum has two isolated equilibrium points: one where the mass M is located vertically at the bottom of the figure (i.e., at 6 o'clock) and the other where the mass is located vertically at the top of the figure (i.e., at 12 o'clock). The model of this pendulum, however, described by Eq. (3.3), has countably infinitely many isolated equilibrium points which are located in R^ at the points (7Tn,0f,n = 0, ± 1 , ± 2 , . . . . • EXAMPLE 3.2. The linear homogeneous system of ordinary differential equations X = A(t)x
(LH)
has a unique equilibrium that is at the origin if A(to) is nonsingular for all to ^ 0. [Refer to Chapter 1 for the definitions of symbols in (LH).] m EXAMPLE 3.3. The linear, autonomous, homogenous system of ordinary differential equations X = Ax
(L)
has a unique equilibrium that is at the origin if and only if A is nonsingular. Otherwise, (L) has nondenumerably many equilibria. [Refer to Chapter 1 for the definitions of symbols in (L).] • EXAMPLE 3.4. Assume that for the autonomous system of ordinary differential equations, X = fix),
{A)
f is continuously differentiable with respect to all of its arguments, and let J{Xe)
dx
(X)
where dfldx denotes the n X « Jacobian matrix defined by dx
dXj
If f{Xe) = 0 and J(Xe) is nonsingular, then Xe is an isolated equilibrium of (A). EXAMPLE 3.5. The system of ordinary differential equations given by xi = k + sin(xi + X2) + xi X2 = k + sin(xi + X2) — xi, with k> I, has no equilibrium points at all.
6.4 QUALITATIVE CHARACTERIZATIONS OF AN EQUILIBRIUM In this section we consider several qualitative characterizations that are of fundamental importance in systems theory. These characterizations are concerned with various
Stability
448 Linear Systems
types of stability properties of an equilibrium and are referred to in the literature as Lyapunov stability. Throughout this section, we consider systems of equations (£"), X = f(t, x),
(E)
and we assume that (E) possesses an isolated equilibrium at the origin. We thus have fit, 0) = 0 for all t > 0. DEFINITION 4.1. The equilibrium ;c = 0 of (E) is said to be stable if for every e > 0 and any ^o ^ R^ there exists a 8(6, to) > 0 such that ||(/)(^, ^0, -^0)11 < ^
for all t > to
(4.1)
whenever Ikoll < 5(6, to).
(4.2)
In Definition 4.1, || • || denotes any one of the equivalent norms on R^, and (as in Chapters 1 and 2) (/)(^, to, XQ) denotes the solution of (E) with initial condition XQ at initial time to. The notation 8(e, to) indicates that 8 depends on the choice of to and e. If in particular it is true that 8 is independent of ^o, i-^-, S = 8(e), then the equilibrium x = 0 of (E) is said to be uniformly stable. In words, Definition 4.1 states that by choosing the initial points in a sufficiently small spherical neighborhood, when the equilibrium x = 0 of (E) is stable, we can force the graph of the solution for t ^ toto lie entirely inside a given cylinder. This is depicted in Fig. 6.1 for the case x G R^. We note that if the equilibrium x = 0 of (E) satisfies condition (4.1) for a single initial condition ^o when (4.2) is true, then it will also satisfy this condition at every initial time t^ > to, where a different value of 8 may be required. To see this, we note that the solution (l)(t, to, xo) determines a mapping g of B(8(€, to)) (at t = to) onto g(B{8(€, to))) (at t ^ t' > to) that contains the origin by assigning for every Xo E B{8{e, to)) one and only one xo = (p(t\ to, xo) E g(B(8(€, to)). By reversing time, cj) determines a mapping of g(B(8(6, to)) onto B(8(e, to)), which is the inverse of ^, denoted by ^~^ Since cf) is continuous with respect to t, to, and xo, then g and g~^ are also continuous. Since 5(S(6, ^o)) is a neighborhood (an open set), then so is g(B(8(e, to)) (refer, e.g., to [16], pp. 320-321). This neighborhood contains in its interior a spherical neighborhood centered at the origin and with a radius §'. If we choose X'Q E B(8'), then (4.1) implies that \\(t)(t, t', X'Q)\\ < e for all t > t^. This argu-
•V"
\^
FIGURE 6.1 Stability of an equilibrium
.•
ment shows that in Definition 4.1 we could have chosen without loss of generality the single value to = 0 in (4.1) and (4.2).
449 CHAPTER 6:
DEFINITION 4.2. The equilibrium x = 0 of (E) is said to be asymptotically stable if (i) it is stable, (ii) for every ^ ^ 0 there exists an i7(^) > 0 such that lim?_oo (l>(t, to, xo) = 0 whenever ||xo|| < rj. m The set of all XQ G J^" such that (pit, to, XQ) ^ 0 as ^ -> oo for some ^o — 0 is called the domain of attraction of the equilibrium x = 0 of (E) (at to). Also, if for (E) condition (ii) is true, then the equilibrium x = 0 is said to be attractive (at to). DEFINITION4.3. The equilibrium x = 0 of (E) is said to be uniformly asymptotically stable if (i) it is uniformly stable, (ii) there is a 5o > 0 such that for every e > 0 and for any to G R'^, there exists a r(e) > 0, independent of to, such that \\(t)(t, to, xo)\\ < e for all ? > ^ + T{e) whenever ||xo|| < 5o. • Condition (ii) in Definition 4.3 can be paraphrased by saying that there exists a So > 0 such that lim (j){t + to, to, xo) = 0 uniformly in (^o, xo) for ^ ^ 0 and for ||xoi| < SQ. In words, this condition states that by choosing the initial points xo in a sufficiently small spherical neighborhood at t = to, WQ can force the graph of the solution to lie inside a given cylinder for all t > to -^ Tie). This is depicted in Fig. 6.2 for the case x E R^. In linear systems theory, we are especially interested in the following special case of uniform asymptotic stability.
^0 + m) FIGURE 6.2 Attractivity of an equilibrium DEFINITION 4.4. The equilibrium x = 0 of (E) is exponentially stable if there exists an a > 0, and for every e > 0, there exists a 6(e) > 0, such that U(t, to, xo)\\ < e^-"^'-^o)
for all t > ^o
whenever \\xo\\ < 6(e) and ^ > 0. Figure 6.3 shows the behavior of a solution in the vicinity of an exponentially stable equilibrium x = 0.
•
Stability
450 Linear Systems
U(o K^
^ee-«(f-^o)
^ ^ / ^ #/ •o.^
~"~""~~---
r 1 1 1
L'^"'
^''"' y
•< -^^ V\
^
\\ \_ee-a(f-g
FIGURE 6.3 An exponentially stable equilibrium DEFINITION 4.5. The equilibrium ;t = 0 of {E) is unstable if it is not stable. In this case, there exists a ^ > 0, an e > 0, and a sequence X;„ ^ 0 of initial points and a sequence {r^} such that |(/)(^o + ^m, ^o. -^m)!! — ^ for all m, ^^ — 0.
•
If X = 0 is an unstable equilibrium of (£"), then it still can happen that all the solutions tend to zero with increasing t. This indicates that instability and attractivity of an equilibrium are compatible concepts. We note that the equilibrium x = 0 of {E) is necessarily unstable if every neighborhood of the origin contains initial conditions corresponding to unbounded solutions (i.e., solutions whose norm grows to infinity on a sequence tm-^ °°)- However, it can happen that a system (E) with unstable equilibrium x = 0 may have only bounded solutions. The concepts that we have considered thus far pertain to local properties of an equilibrium. In the following, we consider global characterizations of an equilibrium. DEFINITION 4.6. The equilibrium x = 0 of {E) is asymptotically stable in the large if it is stable and if every solution of {E) tends to zero as r -> oo. • When the equilibrium x = 0 of (E) is asymptotically stable in the large, its domain of attraction is all of R^. Note that in this case, x = 0 is the only equilibrium of(£). DEFINITION 4.7. The equilibrium x = 0 of iE) is uniformly asymptotically stable in the large if (i) it is uniformly stable, (ii) for any a > 0 and any 6 > 0, and to G R'^, there exists T(e, a) > 0, independent of to, such that if ||xo|| < a, then \\(f)(t, to, xo)|| < e for all t ^ to + T{e, a). • DEFINITION 4.8. The equilibrium x = 0 of {E) is exponentially stable in the large if there exists a > 0 and for any /3 > 0, there exists ^(j8) > 0 such that U{t, to, xo)|| < k(P)\\xo\\e-''^'-''^
for all t > to
whenever II xo 11 < (3.
•
We conclude this section with a few specific cases. The scalar differential equation X = 0
(4.3)
has for any initial condition x(0) = XQ the solution
(4.4)
has for every x(0) == XQ the solution (/)(r, 0, XQ) = xoe^^andx = 0 is the only equilibrium of (4.4). If (3 > 0, this equilibrium is unstable, and when a < 0, this equilibrium is exponentially stable in the large. The scalar differential equation
(4.5)
i = {j=^} has for every x(to) = XQ, ^O ^ 0, a unique solution of the form 0(r, to, xo) - (1 + ^o)^o I ^-^Tj j
(4.6)
and X = 0 is the only equilibrium of (4.5). This equilibrium is uniformly stable and asymptotically stable in the large, but it is not uniformly asymptotically stable. As mentioned earlier, a system X = fit, X)
(E)
can have all solutions approaching an equilibrium, say, x = 0, without this equilibrium being asymptotically stable. An example of this type of behavior is given by the nonlinear system of equations x\{X2 — Xi) + x\
Xi
{x\ -h xl)[\ + {x\ + X^)2] x]{x2 - 2xi)
^2
=
{x\ + xl)[\+{x\
+ xl)^]
For a detailed discussion of this system, refer to Hahn [7], cited at the end of this chapter. Before proceeding any further, a few comments are in order concerning the reasons for considering equilibria and their stability properties as well as other types of stability that we will encounter. To this end we consider linear time-varying systems described by the equations X - A{t)x + B{t)u
(4.7a)
y = C(t)x + D(t)u
(4.7b)
and linear time-invariant systems given by X = Ax + Bu
(4.8a)
y = Cx + Du,
(4.8b)
where all symbols in (4.7) and (4.8) are defined as in Eqs. (6.1) and (6.8) of Chapter 2, respectively. The usual qualitative analysis of such systems involves two concepts, internal stability and input-output stability.
451 CHAPTER 6:
Stability
452 Linear Systems
In the case of internal stability, the output equations (4.7b) and (4.8b) play no ^^^^ whatsoever, the system input u is assumed to be identically zero, and the focus of the analysis is concerned with the qualitative behavior of the solutions of linear time-varying systems X - A{t)x
(LH)
or linear time-invariant systems X = Ax
(L)
near the equilibrium x = 0. This is accomplished by making use of the various types of Lyapunov stability concepts introduced in this section. In other words, internal stability of systems (4.7) and (4.8) concerns the Lyapunov stability of the equilibrium X = 0 of systems (LH) and (L), respectively. In the case of input-output stability, we view systems as operators determined by (4.7) or (4.8) that relate outputs y to inputs u and the focus of the analysis is concerned with qualitative relations between system inputs and system outputs. We will address this type of stability in Section 9 of this chapter.
6.5 LYAPUNOV STABILITY OF LINEAR SYSTEMS In this section we first study the stability properties of the equilibrium x = 0 of linear autonomous homogeneous systems X = Ax,
t^
0,
(L)
and linear homogeneous systems X = A{t)x,
t> tQ>0,
{LH)
where A{t) is assumed to be continuous. Recall that x = 0 is always an equilibrium of (L) and {LH) and that x = 0 is the only equilibrium of {LH) if A{t) is nonsingular for all t>Q. Recall also that the solution of {LH) for x(^) = XQ is of the form ^{t, to, xo) - ^{t, to)xo,
t > to,
where O denotes the state transition matrix of A{t) and that the solution of (L) for x{to) = Xo is given by (l){t, to, Xo) = ^{t,
to)Xo = ^{t
- to, 0)X0
^ 0(r - to)xo = e^^'-'^^xo, where in the preceding equation, a slight abuse of notation has been used. We first consider some of the basic properties of system {LH). THEOREMS.1. The equilibrium jc = 0 of {LH) is stable if and only if the solutions of {LH) are bounded, i.e., if and only if sup||0(r,^)||^
k{to)<^,
where ||0(r, ro)|| denotes the matrix norm induced by the vector norm used on /?", and k{tQ) denotes a constant that may depend on the choice of to.
Proof, Assume that the equiUbrium x = 0 of {LH) is stable. Then for any ^o — 0 and for e = 1 there is a 5 = 5(1, ^ ) > 0 such that ||(^(r, t^, XQ% < 1 for all t > t^ and all XQ with ||xo|| ^ 6. In this case c^(t, to){xo8)
U{t, to, xo)\\ = \\^(h to)xo\\ =
Ikoll
w 8
Ikoll
<
for all ;co ^ 0 and all t ^ to. Using the definition of matrix norm [refer to (10.17) in Chapter 1] it follows that \\^(tJo)\\^S-\
t^to.
We have proved that if the equilibrium x = 0 of {LH) is stable, then the solutions of {LH) are bounded. Conversely, suppose that all solutions (^{t, to, xo) = 0(r, ^o)-^o are bounded. Let {ei,..., en} denote the natural basis for ^-space and let ||(/)(^, to, ej)\\ < (Sj for all t > ^oThen for any vector xo = X'/=i <^7^; we have that U{t. to, xo)\\ = y^
aj(f){t,to,ej)
7=1
7=1
< (max jS;)X 1^7-1 ^
k\\xo\\
;=i
for some constant ^ > 0 for ^ > ^o- For given e > 0, we choose 8 = elk. Thus, if ||xo|| < 6, then ||(^(^, to, xo)|| < ^||^o|| < ^ for all t > to- We have proved that if the solutions of {LH) are bounded, then the equilibrium x = 0 of {LH) is stable. • THEOREM 5.2. The equilibrium x = 0 of {LH) is uniformly stable if and only if ko
sup ^(^o) = sup(sup||0(r,
The proof of this theorem is very similar to the proof of Theorem 5.1 and is left as an exercise. EXAMPLES.1. We consider the system given by 0
(5.1)
e~'
with x{0) = Xo. We transform (5.1) by means of the relation x = Py, where P =
1 1 0 1
p i
=
and obtain the equivalent system Ji LJ2j
(5.2)
0
with y{0) = yo = P~^xo. System (5.2) has the solution (//(r, 0, jo) = "^{t, 0)}^o, where ^1/2(1-^-^0
^(r, 0)
0
0 g(l-.-0
Thesolutionfor(5.1)isobtainedas(/>(r,0, Xo) = 0(r, 0)xo, where ^(r, 0) = From this we obtain for ^o ^ 0, 4){t, to, xo) = ^{t, to)xo, where
P'^{t,0)P~\
453 CHAPTER 6
Stability
454
^(t, to)
0
Je~'0-e-^)
Linear Systems Now ^l/2e"2^0
lim^itJo) =
^e~^0 __
^l/le'^^O
(5.3)
0
We conclude that limy^^oolimr^oo||4)(^, ^o)|| < oo, and therefore (since ||(|>(r, ro)|| ^
y
Xlj=i\ct>ij(tJo)\^ < Xlj=i\cl>ij(^>to)\l that sup,^^^(sup,^J^(tJo)\\) < ^. Therefore, the equihbrium x = 0 of system (5.1) is stable by Theorem 5.1 and uniformly stable by Theorem 5.2. • THEOREM 5.3. The following statements are equivalent. (i) The equilibrium x = 0 of (LH) is asymptotically stable, (ii) The equilibrium x = 0 of (LH) is asymptotically stable in the large, (iii) lim,_oo||c|>(r,ro)|| = 0. Proof, Assume that statement (i) is true. Then there is an 17(^0) > 0 such that when ||xo|| ^ vOo)^ then (^(^, to, xo) -> 0 as ? -> 00. But then we have for any xo ^ 0 that (f){t,tQ,Xo) = Mt,tQ,y]{tQ)
•^0
0
Ikoli
as r —> CO. It follows that statement (ii) is true. Next, assume that statement (ii) is true and fix ^0 ^ 0. For any 6 > 0 there must exist a Tie) > 0 such that for all t ^ to + T(e) we have that \\(l)(t, to, xo)\\ = \\^(t, to)xo\\ < e. To see this, let {e^ . . . , ^„} be the natural basis for R^. Thus, for some fixed constant ^ > 0, if xo = ( a i , . . . , a«)^ and if ||xo|| ^ 1, then xo = 2 " = i (^jej and 2y==i \o^j\ — k. For each J there is a Tj{e) such that \\<^{t, to)ej\\ < elk and ^ > ^0 + Tj{e). Define T{e) = max {Tj{e) : j = I,.. .,n}. For ||xo|| ^ 1 and t ^ to + T(e), we have that
ll^(^,^)^o||
^aj<^(tJo)ej 7=1
1'"^-
By the definition of the matrix norm [see (10.17) of Chapter 1], this means that ||0(r, ^)|| < 6 for f > ^0 + T(e). Therefore, statement (iii) is true. Finally, assume that statement (iii) is true. Then \\^(t, to)\\ is bounded in t for all t > to. By Theorem 5.1, the equihbrium x = 0 is stable. To prove asymptotic stability, fix ro > 0 and e > 0. If ||xo|| < r](to) = 1, then Uit, to, xo)|| < ||^(^, ^)|| ||xo|| ^ 0 as t -^ CO, Therefore, statement (i) is true. This completes the proof. • EXAMPLE 5.2. The equilibrium x = 0 of system (5.1) given in Example 5.1 is stable but it is not asymptotically stable since lim^^oo ||^(^, ^)|| 7^ 0 [see Eq. (5.3)]. • EXAMPLE5.3. The solution of the system -e^'x,
x(to) = Xo
(5.4)
is (/)(?, to, Xo) = ^(t, to)xo, where ^(t, to) = e',(l/2)(e2^0-e20 Since lim^-^ 00 ^(^, ^0) = 0, it follows that the equilibrium x = Oof system (5.4) is asymptotically stable (in the large). • THEOREM 5.4. The equilibrium x = 0 of ilM) is uniformly asymptotically stable if and only if it is exponentially stable.
Proof, The exponential stability of the equilibrium x = 0 implies the uniform asymptotic stability of the equilibrium x = 0 of systems {E) in general, and hence, for systems (L//) in particular. Conversely, assume that the equilibrium x = 0 of {LH) is uniformly asymptotically stable. Then there is a 6 > 0 and a 7 > 0 such that if ||xo|| ^ 5, then \\^{t + to + T, to)xo\\ < ^ for all t, to > 0. This implies that ||cD(/ + /o + r,ro)||< i
ifr,ro^O.
(5.5)
From Theorem 3.6 (iii) of Chapter 2 we have that (^(t, r) = <[>(r, o-)0(a-, r) for any t, cr, and T. Therefore, \\^(t + to + 2T, to)\\ = \\^(t + to + 2Tj + to + T)^{t + to + T, to)\\ ^ |, in view of (5.5). By induction, we obtain for tJo — O that \\^(t + to + nT, to)\\ ^ T\ Now let a = (\n2)/T. Then (5.6) implies that for 0 < ^ < T we have that Uit + to + nT, to, xo)\\ < 2||xo||2-(«+i> = 2||xo|k-"^"^^^^ < 2||;co|k-"^^^'^^>, which proves the result. EXAMPLE 5.4. For system (5.4) given in Example 5.3 we have
(5.6)
U(tJo.xo)\\ = |c/>(Uo,xo)| = \xoe^"^^^'''e-^"'^^''\ < \xo\e^"^^'"''e-
r > ?o > 0,
since e^^ > 2t. Therefore, the equilibrium x = 0 of system (5.4) is uniformly asymptotically stable in the large, and exponentially stable in the large. • Even though the preceding results require know^ledge of the state transition matrix $(r, ro) of {LH), they are quite useful in the qualitative analysis of linear systems. In view of the above results, we can state the following equivalent definitions. The equilibrium x = 0 of {LH) is stable if and only if for any ^ ^ 0 there exists a finite positive constant y = y{to) (which in general depends on ^o) such that for any XQ, the corresponding solution satisfies the inequality U{tJo.xo)\\^
y{to)\\xol
t^
to.
Also, the equilibrium x = 0 of {LH) is uniformly stable if and only if there exists a finite positive constant y (independent of to), such that for any to and JCQ, the corresponding solution satisfies the inequality
U{t,to,xo)\\^
rllxoll,
t^ to.
Furthermore, in view of the above results, if the equilibrium x = 0 of {LH) is asymptotically stable, then in fact it must be globally asymptotically stable, and if it is uniformly asymptotically stable in the large, then in fact it must be exponentially stable (in the large). In this case there exist finite constants y > 1 and A > 0 such that \mto>xo)\\^ye-^^'-'^^\\xo\\ for r > /o > 0 and xo E R"". The results in the next theorem, which are important in their own right, will be required later.
455 CHAPTER 6 :
Stability
456 Linear Systems
THEOREMS.5. Let A be bounded on ( - oo, oo). Then any one of the following statements ^^ equivalent to the exponential stability of the origin x = 0 of {LH):
(i) \;^mt,to)\?dt^ cifox^wh^to. (ii) \;^mtjo)\\dt^ c2fox^\\h^ to.
(iii)
\;^\\^{h,T)fdT^C^foX^\\h^to.
(iv) \;^mh,T)\\dT^ C,for2i\\h^to. (The constants ci are independent of ^ or h, and | • | denotes a matrix norm induced by any one of the equivalent vector norms on /?".) Proof, If the equilibrium x = 0 of {LH) is exponentially stable, then there exist constants y > 0, A > 0 (independent of to) such that \\^{t, ro)|| ^ y^-^(^-^o), t > t^. Substituting this estimate into (i) to (iv) and evaluating the integrals yields ci = C3 = y^l{2K) andc2 = C4 = y/A. Therefore, if the equilibrium X = 0 of (L//) is exponentially stable, then the integrals (i) to (iv) are bounded. We now prove the converse statements by considering each case individually. (a) Assume that the bound c\ in (i) exists. Since A is bounded, there exists an a > 0 such that ||ci>(Uo)|| = ||A(Ocl>(Uo)||^ ||A(0||||^a^)|| < a\\^{t, to)\\, t > ^0. Therefore, ||cD(ri, ^o)^^(^b ^0) - /||
[6(r, tof^it, to) + ^{t, tof(^{t, to)] dt\\ to
' \\^{t, to)^^(t, to) + (D(^, to)^i(t, to)\\dt
\\\[A{t)^(t, to)n mt, to)\\ + mt, to)^ IIA(O^(^, to)\\}dt rh
< 2a
||0(r, to)fdt
< 2aci,
t > to.
J to
Using the triangle inequality of norms yields
mti, to)^^(tu to)\\ = mtu to)^^(tu to)-i+1\\ < ||
ti^
This shows that ||^(ri, ^)|| is itself bounded for ti > ^ by a constant Li. To determine an exponential bound for ||0(f, ^o)|| we note that mt,to)fdT= tQ
i'mt,T)^(TJo)fdT J to
mt, r t • mr,
toifdr
to
< Lf f' ||
to)fdr
Jto
^ LW Noting that the integrand on the left side does not depend on r, we obtain {t - to)\\^{t, to)f < L\ci,
t > ^0.
Let ^ = ^ -f- 7 and let T = AL\ci. Then
mto + Tjo)\\^
i
to.
457
Repeating the above procedure, we obtain
CHAPTER 6 :
110^0 +2r,ro)N ^ 5
Stability
and in general we have ||(^, ^)|| is exponentially bounded. (b) As in (a), but using ^{t\, to) instead of 4>(ri, to)^(^(ti, to), and using the inequality ||i>(r, ^)|| < a||0(f, ^o)||, we obtain llcD^i, ro) - / | | =
\\r^{tJo)dt\\ rh
f'||(i>(Uo)||^^^c.f'||0(Uo)||^^ Jto
JtQ
< Q;C2,
h > ho-
using the inequality property of norms, we have ||c|>(ri, ^0)11 = ||^(ri, to)-I
+ I\\ ^ mh,
to) - I\\ + ||/||
< ac2 + 1. Using this estimate, we now obtain llcDfe to)fdt
< (ac2 + 1) f'' ||Oa to)\\dt
JtQ
JtQ
< (aC2 + l)C2.
This shows that (ii) implies (i), and therefore, it follows that ||0(/, /o)|| is exponentially bounded. (c) The proof of this part follows the proof of (i) with some modifications. Since A is bounded there exists a > 0 such that ||A(0|| ^ « for all t. From Exercise 2.35 in Chapter 2, we have ^-^(t, T) = -(D(?,
dr
T)A(T),
and therefore. ^^(t,T)
= ||-OaT)A(T)||<
a\\^(t,T)\
Then
Mtu rf ^{ti,
•^(ti,to)'^^(hJo)\\
T) + ^(^1, rf
^{h,
T) dr
As in (a), we now obtain ||/ - ^(^1, tof^ty,
to)\\ ^ 2ac3,
ti > to.
Using the triangle inequality, we obtain
mt,, to)^^{ti, to)\\ ^ mtx, to)^^{t,, to) -1\\ + ii/ii <
2Q:C3 +
1.
This shows that ||0(^, ^)|| is bounded for all ^ > ro by a constant L3 that is independent of rand ^0-
458
To obtain an exponential bound we proceed similarly as in (a) to obtain
Linear Systems
u ^ ,.\\^u.
..M|2 ^ f'licT... ..M|2 J to
to
J to
< L 3n
||cD(ri,T)|pjT
Therefore, we have from which we obtain, letting T = 4L3C3, 11^^0 + 7,^0)11^ h and more generally, we obtain mto + nT,to)\\^{kJ^ Again, as in part (a), we conclude that O is exponentially bounded. (d) Assume that (iv) is satisfied. Then as in (b), we can show that (iv) implies (iii), which in turn implies that is exponentially bounded. We obtain 110(^1, ^o) - /|| < ac4,
t > to,
where a is defined as before and C4 is given in (iv). From this we can conclude that \\^(ti, to)\\ is bounded for all ^1 > ^0 by a:c4 + 1. This in turn yields the inequality ['' \\^(t, rtdr
< (ac4 + 1) f' mtu r)\\dT
Jto
JtQ
< (ac4 + 1)C4.
From this it follows that (iv) implies (iii). This concludes the proof of the theorem.
•
We now turn our attention to linear, autonomous, and homogeneous systems given by X = Ax,
r > 0,
(L)
referring to the discussion in Subsection 2.4A [refer to Eqs. (4.11) to (4.27) in Chapter 2] concerning the use of the Jordan canonical form to compute exp (At). We let J = P~^AP and defi[ne x = Py. Then (L) yields y = p-^APy
= Jy.
(5.7)
It is easily verified (the reader is asked to do so in the Exercises section) that the equilibrium jc = 0 of (L) is stable (resp., asymptotically stable or unstable) if and only if y = 0 of (5.7) is stable (resp., asymptotically stable or unstable). In view of this, we can assume without loss of generality that the matrix A in (L) is in Jordan canonical form, given by A = diag [Jo> Ji,' where
/Q = diag [Ai,..., A^]
for the Jordan blocks Ji,..
.,Js.
and
-, Js\y Jk = )^k+iU + M
As in (4.21), (4.22), (4.26) and (4.27) of Chapter 2, we have
Stability
^ o^st
0 where
CHAPTER 6 :
0
(?Jof
oAt
459
„Atn e"''']
e-"" = diagle"",...,
(5.8)
tm-i
and
^•fit =
g^k+it
1
t
0
1
t
0
0
0
2
•••
(«,-l)! {rii - 2)!
(5.9)
1
for / = 1 , . . . , s. Now suppose that Re kt < j8 for all / = 1 , . . . , k. Then it is clear that \imt-,oo (ll^-^o^ll/^^O < °°^ where ||^^°^|| is the matrix norm induced by one of the equivalent vector norms defined on R^. We write this as \\e-^^% = €(e^^). Similarly, if jS = Re Xk+i, then for any e > 0 we have that |^"^^^|| = ©(r^^-^^^O = €(e^^^^^'). From the foregoing it is now clear that \\e^^\\ ^ K for some ^ > 0 if and only if all eigenvalues of A have nonpositive real parts, and the eigenvalues with zero real part occur in the Jordan form only in JQ and not in any of the Jordan blocks JI, 1 < / < 5-. Hence, by Theorems 5.1 and 5.2, the equilibrium x = 0 of (L) is under these conditions stable, in fact uniformly stable. Now suppose that all eigenvalues of A have negative real parts. From the preceding discussion it is clear that there is a constant ^ > 0 and an a > 0 such that ll^^^ll < Ke-""', and therefore, Uit, to, xo)|| < Ke-''^'-'^^xo\\ for alW > ro > 0 and for all xo ^ R'^. It follows that the equilibrium x = 0 is uniformly asymptotically stable in the large, in fact exponentially stable in the large. Conversely, assume that there is an eigenvalue A/ with nonnegative real part. Then either one term in (5.8) does not tend to zero, or else a term in (5.9) is unbounded at ^ ^ oo. In either case, ^^^x(0) will not tend to zero when the initial condition x(0) = XQ is properly chosen. Hence, the equilibrium x = 0 of (L) cannot be asymptotically stable (and hence, it cannot be exponentially stable). Summarizing the above, we have proved the following result. THEOREM 5.6. The equihbrium x = Oof (L) is stable, in fact uniformly stable, if and only if all eigenvalues of A have nonpositive real parts, and every eigenvalue with zero real part has an associated Jordan block of order one. The equilibrium x = 0 of (L) is uniformly asymptotically stable in the large, in fact exponentially stable in the large, if and only if all eigenvalues of A have negative real parts. • A direct consequence of the above result is that the equilibrium x = 0 of (L) is unstable if and only if at least one of the eigenvalues of A has either positive real part or has zero real part that is associated with a Jordan block of order greater than one. At this point, it may be appropriate to take note of certain conventions concerning matrices that are used in the literature. It should be noted that some of these are
460 Linear Systems
not entirely consistent with the terminology used in Theorem 5.6. Specifically, a real nXn matrix A is called stable or a Hurwitz matrix if all its eigenvalues have negative real parts. If at least one of the eigenvalues has positive real part, then A is called unstable. A matrix A, which is neither stable nor unstable, is called critical, and the eigenvalues with zero real parts are called critical eigenvalues. We conclude our discussion concerning the stability of (L) by noting that the results given above can also be obtained by directly using the facts established in Subsection 2.4C, concerning modes and asymptotic behavior of time-invariant systems. EXAMPLE 5.5. We consider the system (L) with 0 1 -1 0 The eigenvalues of A are Ai, A2 = ±j. According to Theorem 5.6, the equilibrium x = 0 of this system is stable. This can also be verified by computing the solution of this system for a given set of initial data x(0)^ = (xi(0), X2(0)), A
(j)i(t,0,xo) = xi(0)cosr + X2(0)sin^ 4>2(t, 0, XQ) = -xi(0)smt + jC2(0)cosr, r > 0, and then applying Definition 4.1.
•
EXAMPLE 5.6. We consider the system (L) with
ro n
A = 0 0 The eigenvalues of A are Ai = 0, A2 = 0. According to Theorem 5.6, the equilibrium X = 0 of this system is unstable. This can also be verified by computing the solution of this system for a given set of initial data x(0)^ = (xi(0), ^2(0)), (l)l(t, 0, Xo) = Xi(0) + X2(0)t (f>2it, 0, Xo) = X2(0), ^ > 0, and then applying Definition 4.5. (Note that in this example, the entire xi-axis consists of equilibria.) • EXAMPLE 5.7. We consider the system (L) with ^ [2.8 9.6] [9.6 -2.8j* The eigenvalues of A are Ai, A2 = ±10. According to Theorem 5.6, the equilibrium X = 0 of this system is unstable. • EXAMPLE 5.8. We consider the system (L) with [-1 -2\The eigenvalues of A are Ai, A2 = - 1 , - 2 . According to Theorem 5.6, the equilibrium X = 0 of this system is exponentially stable. • Next, we consider linear periodic systems given by X = A(t)x,
A(t) = A(t + Tl
(P)
where A(t) is a continuous real matrix for all f G (-00, 00). We recall from Section 2.5 of Chapter 2 that if 0(^, ^0) is the state transition matrix for (P), then there exists
a constant nX n matrix R and a nonsingular nX n matrix "^(t, to) such that ^(t, to) = nt, to) exp [R(t - to)l
461 (5.10)
CHAPTER 6:
Stability where
^(? + T, to) = "i^it, to)
for all t E (-00, oo). Now according to the discussion at the end of Section 2.5 of Chapter 2, the change of variables x = '^(t, to)y transforms (P) to the system (5.11)
y = Ry^
where R is given in (5.10). Since ^(t, to) is nonsingular, it is clear that the equilibrium X = 0 of (P) is uniformly stable (resp., uniformly asymptotically stable) if and only if the equilibrium j == 0 of system (5.11) is uniformly stable (resp., uniformly asymptotically stable). Now by applying Theorem 5.6 to system (5.11) we obtain the following result. THEOREM 5.7. The cquiHbrium x = 0 of (P) is uniformly stable if and only if all eigenvalues of the matrix R [given in (5.10)] have nonpositive real parts, and every eigenvalue with zero real part has an associated Jordan block of order one. The equilibrium X = 0 of (P) is uniformly asymptotically stable in the large if and only if all eigenvalues of R have negative real parts. •
6.6 SOME GEOMETRIC AND ALGEBRAIC STABILITY CRITERIA In this section we concern ourselves with nth-order linear homogeneous ordinary differential equations of the form aox^""^ + axx^""-^^ +• -h an-ix^^"^ -h anX = 0,
ao 7^0,
(6.1)
where the coefficients ao,.. .,an are all real numbers. We recall from Chapter 1 that (6.1) is equivalent to the system of first-order ordinary differential equations (6.2)
X = Ax, where in (6.2) A denotes the companion matrix given by
0 0
1 0
0 1
0 0
_ '^"-2
fll
(6.3)
A = On
_a„-i
ao
ao
flo
flO
To determine whether the equilibrium x = 0 of (6.3) is asymptotically stable, it suffices to determine if all the eigenvalues of A have negative real parts, or what amounts to the same thing, if the roots of the polynomial /(A) = aoX"" + (21A"""^ -h • • • + (2^-1 A -h an
(6.4)
all have negative real parts. To see this, we must show that the eigenvalues of A coincide with the roots of the polynomial f(s). This is most easily accomplished by induction. For the first-order case k = 1, we have A = -ajao and therefore det(\I\ - A) ^ A -h aJao, h = I, and so the assertion is true foYk= 1. Next, assume that the assertion is true for k = n - I. Then
462 Linear Systems
-1 A
0 -1
0 0
0 0
0
0
A 02
- 1 A + fli
flO
ao
ao
ao
det{XIn - A) = 0
= Xdet(Xln-\ - Ai) + (-1>n+l
"n
-1 A 0
0 -1 A
0
0
0 0 -1
0 0 0
ao
0 0 where
1 0
0 1
0 0
an-3
ai
ao
flo
Ai =
a„-i
_an-2
flO
and In, In-\ denote the n X /i and {n-
ao \)X {n-
1) identity matrices. Therefore,
det(\In - A) = Xdet(\In-i
- Ai) + — ao = A" + ^ A " - i + A+ — ao ao ao 0
is equivalent to /(A) = 0. Analogously to matrices, we now make the following definitions. An nth-order polynomial /(A) with real coefficients [such as (6.4)] is called stable if all zeros of /(A) have negative real parts, it is called unstable if at least one of the zeros of /(A) has a positive real part, and it is called critical if /(A) is neither stable nor unstable. Also, a stable polynomial is called a Hurwitz polynomial. In view of the above, the stability problem for nth-order differential equations with constant coefficients has now been reduced to a purely algebraic problem of determining whether the zeros of a polynomial [such as (6.4)] all have negative (resp., nonpositive) real parts. In case zeros with vanishing real parts exist, it is further necessary to determine their multiplicity. In the following, we first present some graphical criteria that enable us to determine the stability of a polynomial (6.4), without determining its roots explicitly. These results, which are important in their own right, are then used to arrive at some algebraic criteria to determine the stability of a polynomial (6.4).
A. Some Graphical Criteria In establishing our first result we assume that the polynomial f(s) [given by (6.4)] has p zeros in the right half of the ^--plane and (n - p) zeros in the left half of the ^--plane, and we assume that there shall be no zeros on the imaginary axis. We let C denote the counterclockwise contour formed by a semicircle C with radius r and cen-
tered at the origin, together with its diameter on the imaginary axis, and we choose r so large that the p zeros of f(s) in the right half of the ^-plane lie in the interior of the circle. We now recall from a well-known result from the elementary theory of functions of a complex variable (called Cauchy's integration formula) that P= ^ ' l ^ T T ^ ^ = ^ 8 c In f(sl (6.5) 277-J )c f{s) 27TJ where 5- is a complex variable, j = ^-l, f'(s) denotes the derivative off with respect to s, \(^[f'(s)/f(s)] ds denotes the integral of f'(s)/f(s) along the contour C, and 8c In f(s) represents the increment of In f(s) along the contour C. Let Sq In f(s) denote the increment of ln/(^) along the semicircular arc C with s = re^"^. For s on C we have f{s) = aor'^^^'"^(l + 0(r"^)) and
ln/(^) = Inao + ^ I n r + njcf) + 0(r"^).
Hence, 8c>lnf(s)
=
njl^^'^
= niTJ + ©(r"^). Letting r ^ oo and using (6.5), we conclude that -n^^j + hs—: ]8i In f(s) 277 j \27rj = ^ + 2:^5/In/(^),
(6.6)
where S/ In f(s) denotes the increment of the logarithm of f(s) along the imaginary axis / from -oo to +oo. To determine this increment, we let s = jo) and let f(s) = f(jay) ^ Ri(o)eJ'(^^ ^ U{co) + jVico),
(6.7)
and we consider R, 6 as polar coordinates and U, V as coordinates in the complex plane. As the real parameter co (the real frequency w) ranges from +00 to -00, the points f(jo)) in (6.7) describe the frequency response or frequency plot for f(s). Since/(s-) has real coefficents, we must have 7?(CL)) = R(-a)) and 0((o) = -6(-(o). It therefore suffices to consider the part of the response curve belonging to the positive values of the parameter 0;. It follows from (6.6) that n _ ^(00) _
2
1
~V~ ~ 2
6(00)
(6.8)
where ^(00) is the limit to which the polar angle 6(a)) = tan~^[V(o))/U((o)] of the frequency response diagram tends as co becomes unbounded. Since U(co) and V((o) are polynomials of different degrees, \V(CO)/U((JO)\ will tend to either zero or infinity. In either case, in view of (6.8), ^(00) must be an integral multiple of 77/2. Now when in particular p = 0, then by necessity we have that 6(00) = n(7r/2). This yields the following result. THEOREM 6.1. (LEONHARD-MIKHAILOV STABILITY CRITERION) The polynomial f(s) has only zeros with negative real parts if and only if its frequency response diagram f(jco), 0 < co < 00, passes through exactly n quadrants in the positive sense. •
463 CHAPTER 6:
Stability
464 Linear Systems
(b) Frequency response plot for f{s) unstable (n = 6)
(a) Frequency response plot for f{s) stable (n = 6)
FIGURE 6.4 Frequency response plot for f(s) stable (n
6)
In Fig. 6.4 we depict the frequency response plots of a stable and an unstable polynomial. Next, since f(s) has real coefficients, there are real polynomials f\ and /2 such that fijco)
(6.9)
^ Moj^) 4- j(of2(co^\
If n = 2k, then deg fi(u) = k and deg f2(u) = k - I, and if n = 2k -{- 1, then deg f\{u) = k and deg f2(u) = k. To the zeros uik, i = 1, 2 and k = 1, 2, 3 , . . . of the polynomials fi(u), /2(w), respectively, correspond those values of co^ at which the frequency response diagram intersects the axes. In the case of stability, these values ofcx)^ must be real and increasing, i.e., the zeros ui^ must be positive and alternate. 0 < Wii < U2\ < Ui2 < U22 <
(6.10)
'"
since otherwise the response curve f{jco) will not make the proper number of turns at the appropriate locations. These considerations lead to the interlacing of the roots u\k with the roots U2k, k = 1, 2, 3 , . . . , which is sometimes called the gap and position criterion. In Fig. 6.5 we depict typical situations for stability and instability (in terms of the variables U and V). We summarize the above in the following result. u, V k
u, V k
r
CO
(a) Stable case FIGURE 6.5
(a) Unstable case
THEOREM 6.2. (GAP AND POSITION STABILITY CRITERION) The polynomial
f(s) has only zeros with negative real parts if and only if the zeros of the polynomials defined in (6.9) are real and satisfy the inequalities (6.10). • The frequency response plot f(j(o) is unbounded and hence can never be displayed completely. We therefore frequently make use of the reciprocal frequency response diagram, i.e., the response diagram of the function l/f(jo)). Such a plot approaches zero asymptotically, and in case of stability, it rotates exactly through n quadrants in the negative sense. In applications of Theorem 6.1, the entire plot is not needed. It can be shown that it suffices to consider the interval 0 < w < CDQ^ where (OQ = 1 + 3 max\p\\.
(6.11)
We will not pursue the details concerning the proof of this assertion.
B. Some Algebraic Criteria In the next results, we develop the Routh-Hurwitz criterion, which yields necessary and sufficient conditions for f(s) to be a Hurwitz polynomial. To accomplish this, we will make use of Theorem 6.2. We begin by establishing a set of necessary conditions. THEOREM 6.3. For
f(s) = aos"" + ais'"''^ + • • • + an-is + <2„
(6.12)
to be a Hurwitz polynomial it is necessary that the inequalities ^>0,^>0,...,^>0 Go
(6.13)
ao
UQ
hold. Proof, Let si,.. .,SnhQ the zeros of (6.12), and in particular, let s'j be the real roots and s'l the complex roots. Then f(s) = aoYlis - s'j)Yl(s - s'l) j
k
= aoYlis - s'j)Yl(s^ - {2Res'l)s + |4f). j
k
If all the numbers s'j and Re s'l are negative, then we can obtain only positive coefficients for the powers of s when we multiply the product out to obtain f{s). • Without loss of generality, we assume in the following that ao > 0. In the next result, we will require the Routh array: ao ri = — , ax r^ =
Cll
c\Q = ao,
C20 = <32
C30 = a4,
C40 =
as,...
Cii = ai,
C21 = as,
C31 = as,
C41 =
aj,...
C22 = a^ - r2as,
^332 = ^6 - ^2<37,...
C23 = C31 - r3C32,
C33 = C41 - r3C42, . . .
C12 = a2,
r2fl3,
Ci3 = C21 - r3C22,
cn Cij = Ci+ij-2 - rjCi+ij-i, <^1,/-1 C\n = an
i=l,2,...,
J = 2, 3 , .
465 CHAPTER 6*
Stability
466 Linear Systems
Note that if n = 2m, we have ^
^^_^
^^_^
^
^^_^
.— n
and if n = 2m - 1, we have
The above array terminates after {n—\) steps in case all the numbers ctj are different from zero. The last line defines ci„. In addition to the inequalities (6.13), we shall require the inequalities given by C\\ > 0, Ci2 > 0, . . ., Cin > 0.
(6.14)
THEOREM 6.4. (ROUTH-HURWITZ STABILITY CRITERION) The polynomial
f{s) given in (6.12) is a Hurwitz polynomial if and only if the inequalities (6.13) and (6.14) hold. Proof, First we assume that the degree of f{s) is even by letting n = 2m. We define the polynomials hi(s) = i [ / ( ^ ) + f(-s)l
h2(s) = l [ / ( ^ ) - f(-s)l
(6.15)
Applying the Euclidean algorithm to determine the greatest common divisor of hi (s) and h2(s), we obtain hi(s) = r2sh2(s) — hsis) h2(s) = r!^sh3(s) - h^is)
(6.16)
where the linear factors arising in the division have no constant term and the remainders have been written with negative signs. It is readily verified that the constants r[ in (6.16) are related to the constants r/ in the Routh array by the expression r- = ( - 1)V/. Next, we define a sequence of polynomials given by h2i-i{s) = g2i-i(s^X
h2i{s) ^ sg2i{sh
i = \,,.,,m.
(6.17)
From (6.16) and (6.17) we obtain the recursion formulas given by gii+iiz) = r2iZg2i(z) - g2i-i(z) g2i+2(z) = r2i+ig2i+i(z) - g2i(z).
(6.18)
The first two members of this sequence are given by gi(z) = aoz^ + a2Z^-i +
'"+a2m
giiz) = aiz""-' + a^f"-^ + ••• + a2m-x.
(6.19)
We can readily verify that the above two polynomials agree with the polynomials /i and /2 given in (6.9), except for sign. In fact, we have fi{u) = gi{-u)
i = 1,2.
(6.20)
We are now in a position to construct the Routh array of a polynomial f{s) by utilizing the coefficients of the sequence of polynomials gi{z)- If in the process of doing so we encounter a zero row (i.e., an identically vanishing polynomial, say, gi), then h\ and /z2 [and thus, f{s) and f{-s)\ have a common divisor. In this case f{s) possesses a divisor of the form s^ + a. and is not a Hurwitz polynomial. Next, we assume that the hypotheses of this theorem are satisfied [i.e., (6.13) and (6.14) hold]. Applying definitions, we can readily verify that the numbers r[ have alter-
nating signs, that the signs of the leading coefficients of the polynomials gi, i = 1,2,... are siven bv ^
-^
467 /: CHAPTER 6 :
+, +, - , - , +, +, - , - , • • •,
(6.21)
and that the degrees of these polynomials are given by m,m— \,m — 1, m — 2, m — 2 , . . . , 1, 1, 0.
(6.22)
Next, for fixed z, -oo < z < oo, we consider the sequence of numbers given by gl{z),g2{z\...,g2m{z).
(6.23)
Note that the last term gimiz) is a constant for all z. Let W{z) denote the number of sign changes in this sequence. When z > 0 and is very large, then the signs of the gi in (6.23) correspond to the signs of the leading coefficients of these polynomials. When - z > 0 and is very large, then the signs of gi in (6.23) will alternate. It now follows that the difference W{-^) — W{+^) is always equal to m. Thus, as z varies from -oo to +°o, m sign changes in the sequence (6.23) will disappear. Such a disappearance can occur only at a zero of g\{z), since if z passes through a zero, say, z', of gi{z), \ < i < 2m, no disappearance in the number of sign changes occurs, because by (6.18), sgngi-\{z') ^ sgn gi+i. We conclude that gi{z) has exactly m real zeros, and since by hypothesis all its coefficients are positive, no positive zeros can occur. A similar argument as above shows that g2 has exactly (m - I) negative zeros. Now, since in the sequence (6.23) the largest possible number of disappearances of sign changes occurs (as z is varied from -oo to +oo), a disappearance in sign change must actually occur each time z passes through a zero of ^i. This, however, is possible only if g2 in turn changes sign between every two zeros of ^ i ; otherwise there would be an additional change of signs in the sequence (6.23). It follows that the zeros of g2 separate the zeros of ^i (the zeros of g2 are interlaced with the zeros of ^i). Now, in view of (6.20), the above statement concerning the zeros of ^i and g2 is equivalent to the inequahty (6.10); thus. Theorem 6.2 applies. Therefore, condition (6.14) is sufficient for f{s) to be a Hurwitz polynomial. It is also a necessary condition, since otherwise the count of the sign changes in the sequence (6.23) is too small and the hypotheses of Theorem 6.2 are not satisfied, i.e., either ^i has too few zeros or the polynomial does not satisfy condition (6.10). To complete the proof, we assume next that n = 2m + 1, i.e., n is odd. In this case we interchange the definitions of hi(s) and h2(s) given in (6.16), and we let h2i+l(s) = Sg2i + \(S^\
h2i(s) = g2i(s'^y
The degrees of the polynomials gi formed in this manner are m, m, m - 1 , m - 1 , . . . , 1, 1, 0. Following a similar procedure as before, we show that the polynomials gi(z) and g2(z) each have m negative zeros and that Theorem 6.2 applies. This concludes the proof. • EXAMPLE6.1. We apply the Routh-Hurwitz criterion (Theorem 6.4) to the polynomial f(s) = (s + 2)(s + 1 - j)(s + 1 + j)(s + 1) = / + 5^^ + 10^2 + 10^ + 4.
(6.24)
For this polynomial we form the Routh array and obtain / s^ s^
1 5 i(5-10-1-10) = 8
s'
| ( 8 - 1 0 - 5 - 4 ) = 7.5
s'
7^[(7.5)-4-8-0] = 4
10 10 1(5-4-0) = 4
4 0 0
i(8-0-5-0) =0 0 0
0
Stability
468 Lh^i^Systems
The conditions of Theorem 6.4 are clearly satisfied. Hence, (6.24) is a Hurwitz polynomial. • EXAMPLE6.2. We apply the Routh-Hurwitz criterion (Theorem 6.4) to the polynomial f(s) = (s + 2)(s + 1 - j)(s + 1 + jXs - 1) = / + 3^^ + 2s^ -2s-4.
(6.25)
We note that condition (6.13) is violated, and therefore, the polynomial (6.25) is not a Hurwitz polynomial. Condition (6.14) is also not satisfied. To see this, we form the Routh array for (6.25), given as / I 2 4 s^ 3 -2 0 s^ i(3-2 + 2 ) = f i[3.(-4)-0] = -4 0.
s' i[f •(-2)-(-4)-3] = f s' i[f •(-4)-0] = -4
0
0 •
6.7 THE MATRIX LYAPUNOV EQUATION In Section 6.6 we established a variety of stability results that require explicit knowledge of the solutions of (L) or (LH). We also derived some geometric and algebraic stability criteria for (L) when the matrix A is in companion form that do not require explicit knowledge of solutions, but instead, are deduced directly from the parameters of A. In this section we will develop stability criteria for (L) with arbitrary matrix A. In doing so, we will employ Lyapunov's Second Method (also called Lyapunov's Direct Method) for the case of linear systems (L). This method utilizes auxiliary real-valued functions v(x), called Lyapunov functions, that may be viewed as "generalized energy functions'' or "generalized distance functions'' (from the equilibrium X = 0), and the stability properties are then deduced directly from the properties of v{x) and its time derivative v(x), evaluated along the solutions of (L). A logical choice of Lyapunov function is v(x) = x^x = ||x|p, which represents the square of the Euclidean distance of the state from the equilibrium x = 0 of (L). The stability properties of the equilibrium are then determined by examining the properties of v(x), the time derivative of v(x) along the solutions of (L), X = Ax.
(L)
This derivative can be determined without explicitly solving for the solutions of (L) by noting that v(x) = xF X + x^ X = (Ax)^x + x^(Ax) = Jc^(A^ + A)x. If the matrix A is such that v(x) is negative for all x T^ 0, then it is reasonable to expect that the distance of the state of (L) from x = 0 will decrease with increasing time, and that the state will therefore tend to the equilibrium x = 0 of (L) with increasing time t. It turns out that the Lyapunov function used in the above discussion is not sufficiently flexible. In the following we will employ as a "generalized distance function"
469
the quadratic form given by v(x) = X Px,
(7.1)
P\
where P is a real nXn matrix. The time derivative of v{x) along the solutions of (L) is determined as v{x) = x^Px + x^Px = x^A^Px
+ x^PAx
= x^{A^P + PA)x, V = x^Cx,
I.e.,
(7.2) (7.3)
C = A^P + PA.
where
Note that C is real and C^ = C. The system of equations given in (7.3) is called the Lyapunov Matrix Equation. We recall from Section 6.2 that since P is real and symmetric, all its eigenvalues are real. Also, we recall that P is said to be positive definite (resp., positive semidefinite) if all its eigenvalues are positive (resp., nonnegative), and it is called indefinite if P has eigenvalues of opposite sign. The definitions of negative definite and negative semidefinite (for P) are similarly defined. Furthermore, we recall that thQ function v(x) given in (7.1) is said to be positive definite, positive semidefinite, indefinite, and so forth, if P has the corresponding definiteness properties (refer to Section 6.2). Instead of solving for the eigenvalues of a real symmetric matrix to determine its definiteness properties, there are more efficient and direct methods of accomplishing this. We now digress to discuss some of these. Let G = [gij] be a real nX n matrix (not necessarily symmetric). Referring to Subsection 2.2G, we recall that the minors of G are the matrix itself and the matrix obtained by removing successively a row and a column. The principal minors of G are G itself and the matrices obtained by successively removing an ith row and an ith column, and the leading principal minors of G are G itself and the minors obtained by successively removing the last row and the last column. For example, if G = [gij] G R^^^, then the principal minors are
gn
gn
gu]
821
gii g32
g23
_g3l
gn _g3l
gl3 g33_
gn
>
gl2 g22_
g2l
^33 J g22
g23
[g32 ^33.
[gnl
[g22l
[g33l
The first three matrices above are the leading principal minors of G. On the other hand, the matrix g21
g22
_g3l
g32
is a minor but not a principal minor. The following results, due to Sylvester, allow efficient determination of the definiteness properties of a real, symmetric matrix. PROPOSITION 7.1. (i) A real symmetricmatrix P = [pij] G R''^''is positive definite if and only if the determinants of its leading principal minors are positive, i.e., if and only if
CHAPTER 6:
Stability
470
Pn > 0,
det \Pn
[Pl2
Linear Systems
Pul >0,...,detP>0. Pll]
(ii) A real symmetric matrix P is positive semidefinite if and only if the determinants of all its principal minors are nonnegative. • Still digressing, we consider next the quadratic form v(w) = w^Gw,
G = G^,
where G G R^^^, Referring to Subsection 6.2C [in particular, Eqs. (2.35) and (2.36)], there exists an orthogonal matrix Q such that the matrix P defined by P = Q-^GQ
=
Q^GQ
is diagonal. Therefore, if we let w = Qx, then v(2;c) = v{x) = x^Q^GQx
=
x^Px,
where P is in the form given in Eq. (2.35), i.e..
0 A2
From this, we immediately obtain the following useful result. PROPOSITION 7.2. Let P = P^ G R''^'', let \M(P) and A^(P) denote the largest and smallest eigenvalues of P, respectively, and let || • || denote the Euclidean norm. Then A^(P)||x|p < v(x) = x^Px <
(7.4)
\M(P)\\xf
for all X E R\ Let ci = \m(P) and C2 = ^M(P)' Clearly, v(x) is positive definite if and only if C2 ^ ci > 0, v(x) is positive semidefinite if and only if C2 ^ ci > 0, v(x) is indefinite if and only if C2 > 0, ci < 0, and so forth. We are now in a position to prove several results. THEOREM 7.1. The equilibrium x = 0 of (L) is uniformly stable if there exists a real, symmetric, and positive definite nX n matrix P such that the matrix C given in (7.3) is negative semidefinite. Proof. Along any solution (/)(r, to, XQ) = (f)(t) of (L) with (/>(^, ^o, XQ) = (^(^) have ct>(tfPct>(t)
d xlPxo + I ^(j>{y]fP(i>{r])dr] = x^Pxo +
xo, we
(fy{r]f C(j>{'r])dri
for all t > to > 0. Since P is positive definite and C is negative semidefinite, we have (l>{tYP(fy{t) - xlPxo
< 0
CHAPTER 6: Stability
for all t > to > 0, and there exist C2 ^ ci > 0 such that
ciwmf ^ m^pm < xiPxo < c2iixoip for all ^ > ^ . It follows that
||<^(ONgj Nl for a l l r > ^ > 0 and for any XQ G R^. Therefore, the equilibrium x = 0 of (L) is uniformly stable (refer to Sections 6.4 and 6.5). • EXAMPLE 7 . 1 . For the system given in Example 5.5 we choose P = I, and we compute C = A^P + PA = A ^ + A = 0. According to Theorem 7.1, the equilibrium x = 0 of this system is stable (as expected from Example 5.5). • THEOREM 7.2. The equilibrium x = 0 of (L) is exponentially stable in the large if there exists a real, symmetric, and positive definite nX n matrix P such that the matrix C given in (7.3) is negative definite. Proof, W e let (f){t, to, xo) = cf>(t) denote an arbitrary solution of (L) with (/>(^) = ;co. In view of the hypotheses of the theorem, there exist constants C2 ^ ci > 0 and C3 > C4 > 0 such that
and
-C3\\m\? ^ v(0(O) = m ^ c m ^ -C4\\m\?
for all / > ^ > 0 and for any ;co e /?". Then
v(m) = j^im^pm] ^ (-^Am^pm -|)v(c^(0) for alU > ^ > 0. This implies, after multiplication by the appropriate integrating factor, and integrating from to to t, that V((/>(0) = (f)(tf P(t>{t) < xJPX0^~^'4/c2)a-%) or or
^ m'^Pm ^ C2\\xo\fe-^'^''^^^'-'^^ I \l/2 ||c/>(0||<^ ||xo|k-^^/^^^^4/c2)a-^o),
471
ciWmf
t^to^^.
This inequality holds for all xo E R^ and for any to ^ 0. Therefore, the equilibrium X = 0 of (L) is exponentially stable in the large (refer to Sections 6.4 and 6.5). • In Fig. 6.6 we provide an interpretation of Theorem 7.2 for the two-dimensional case {n ^ 7). The curves Q , called level curves, depict loci where v{x) is constant, i.e., Ci = {x G R^ \ v(x) = x^Px = Ci}, i = 0, 1, 2, 3, When the hypotheses of Theorem 7.2 are satisfied, trajectories determined by (L) penetrate level curves corresponding to decreasing values of c/ as t increases, tending to the origin as t becomes arbitrarily large.
472 Linear Systems C2 = {xeR^:v{x)
= C:i}
Ci =[xeR^:
v{x) = c-]}
\
Co = {xe R^ : v(x) = Co = 0} 0 =
|
C2 = { x € H ^ : v(x) = C2}
C-\ < C2 < C3 • • "
CQ<
to
< f2 < ^3 • • •
-fc
V(X) = C3 '
v{x) = C2 -""^ \/(x) = ci
—^-^2
^^N^^^Vl
^^
- ^ FIGURE 6.6 Asymptotic stability
>^1
EXAMPLE 7.2. For the system given in Example 5.8, we choose
^
n 0 0
0.5
and we compute the matrix C = A^P + PA =
-2 0
0 -2
According to Theorem 7.2, the equilibrium x = 0 of this system is exponentially stable in the large (as expected from Example 5.8). • THEOREM 7.3. The equilibrium x = 0 of (L) is unstable if there exists a real, symmetric nX n matrix P that is either negative definite or indefinite such that the matrix C given in (7.3) is negative definite. Proof, We first assume that P is indefinite. Then P possesses eigenvalues of either sign, and every neighborhood of the origin contains points where the function
v(x) = x^Px
473 CHAPTER 6:
is positive and negative. Consider the neighborhood
Stability
B{e) = {x e /?^ : \\x\\ < e}, where || • || denotes the EucHdean norm, and let G = {x e B{e) : v(x) < 0}. On the boundary of G we have either ||x|| = e or v{x) = 0. In particular, note that the origin x = 0 is on the boundary of G. Now, since the matrix C is negative definite, there exist constants C3 > Q > 0 such that -csll^lp < x^Cx = v(x) < -QII^IP
for all X E /^". Let (t)(t, to, XQ) = (i){t) and let xo = (/>(^) G G. Then v(xo) = -(3 < 0. The solution (f){t) starting at JCQ must leave the set G. To see this, note that as long as (f){t) E G, v((/)(0) ^ - « since v(x) < 0 in G. Let -c = sup {v(x) : x ^ G and v(x) < -fl}. Then c > 0 and v{(f){t)) = v(xo) +
v((t)(s))ds < -(2 -
C(i5
= —a — (t — to)c, t > tQ.
This inequality shows that 4>{t) must escape the set G (in finite time) because v{x) is bounded from below on G. But (f){t) cannot leave G through the surface determined by v{x) = 0 since v((/)(0) ^ - a . Hence, it must leave G through the sphere determined by ||x|| = 6. Since the above argument holds for arbitrarily small 6 > 0, it follows that the origin x = 0 of (L) is unstable. Next, we assume that P is negative definite. Then G as defined is all of B(e), The proof proceeds as above. • The proof of Theorem 7.3 shows that for e > 0 sufficiently small when P is negative definite, all solutions (pit) of (L) v^ith initial conditions XQ E B(e) w^ill tend aw^ay from the origin. This constitutes a severe case of instability, called complete instability. EXAMPLE 7.3. For the system given in Example 5.7, we choose 0.28 -0.961 P= -0.96 0.28 and we compute the matrix C = A^P + PA =
-20 0
0 -20
The eigenvalues of P are ±1. According to Theorem 7.3, the equilibrium x = 0 of this system is unstable (as expected from Example 5.7). • In applying the results derived thus far in this section, we start by choosing (guessing) a matrix P having certain desired properties. Next, we solve for the matrix C, using Eq. (7.3). If C possesses certain desired properties (i.e., it is negative definite), we draw appropriate conclusions by applying one of the preceding theorems of this section; if not, we need to choose another matrix P. This points to the
474 Linear Systems
principal shortcoming of Lyapunov's Direct Method, when appHed to general systerns. However, in the special case of linear systems described by (L), it is possible to construct Lyapunov functions of the form v(x) = x^Px in a systematic manner. In doing so, one first chooses the matrix C in (7.3) (having desired properties), and then one solves (7.3) for P. Conclusions are then drawn by applying the appropriate results of this section. In applying this construction procedure, we need to know conditions under which (7.3) possesses a (unique) solution P for a given C. We will address this topic next. We consider the quadratic form v(x) = x^Px,
P = P^,
(7.5)
and the time derivative of v(x) along the solutions of (L), given by v(x) = x^Cx, where
C = C^,
(7.6)
C = A^P + PA
(1.1)
and where all symbols are as defined in (7.1) to (7.3). Our objective is to determine the as yet unknown matrix P in such a way that v(x) becomes a preassigned negative definite quadratic form, i.e., in such a way that C is a preassigned negative definite matrix. Equation (7.7) constitutes a system of n(n + l)/2 linear equations. We need to determine under what conditions we can solve for the n(n + l)/2 elements, pik, given C and A. To this end, we choose a similarity transformation Q such that QAQ-^ = A,
(7.8)
A = Q-^AQ,
(7.9)
or equivalently,
where A is similar to A and 2 is a real nX n nonsingular matrix. From (7.9) and (7.7) we obtain (AfiQ-^PQ-^ + (Q'VPQ'^A or
(A/Q ^QA = C,
P = (Q~VPQ~\
= (Q-VCQ~^ C = {Q-^fCQ'K
(7.10) (7.11)
In (7.11), P and C are subjected to a congruence transformation and P and C have the same definiteness properties as P and C, respectively. Since every real nX n matrix can be triangularized (refer to Subsection 2.2L), we can choose Q in such a fashion that A == [dij] is triangular, i.e., dtj = 0 for / > j . Note that in this case the eigenvalues of A, A i , . . . , A„, appear in the main diagonal of A. To simplify our notation, we rewrite (7.11) in the form (7.7) by dropping the bars, i.e., A^P^-PA
= C,
C = C^,
(7.12)
2indwe assume that A = [aij] has been triangularized, i.e., atj = 0 f o r / > j . Since the eigenvalues A i , . . . , A„ appear in the diagonal of A, we can rewrite (7.12) as 2Xipn
= cii
^2iPii + (Ai + X2)pn
= ci2
(7.13)
Since this system of equations is triangular, and since its determinant is equal to 2^Ar • -A^ n ( A / + A/),
(7.14)
i^j
CHAPTER6:
Stability
the matrix P can be determined (uniquely) if and only if its determinant is not zero. This is true when all eigenvalues of A are nonzero and no two of them are such that \i + Xj = 0. This condition is not affected by a similarity transformation and is therefore also valid for the original system of equations (7.7). We summarize the above discussion in the following lemma. LEMMA 7.1. Let A E /?"^" and let Ai,..., A„ denote the (not necessarily distinct) eigenvalues of A. Then (7.12) has a unique solution for P corresponding to each C E /?«><« if and only if A, 7^ a A, + Ay 7^ 0
475
for all /, ;.
(7.15) •
To construct v(x), we must still check the definiteness of P. This can be done in a purely algebraic way; however, in the present case it is much easier to apply the results of this section and argue as follows: 1. If all eigenvalues of A have negative real parts [or equivalently, if the equilibrium jc = 0 of (L) is exponentially stable in the large], and if C in (7.7) is negative definite, then P = P^ must be positive definite. To prove this assertion, we choose for (L) the function v given in (7.5) with v along the solutions of (L) given by (7.6) and (7.7). For purposes of contradiction we assume that P is not positive definite. Then there exists XQ T^ 0 such that V(JCO) = :^o ^^o — 0. For the solution (pit) with 4>(to) = XQ, V((/)(0) is monotone decreasing with increasing t, since v((/)(0) — 0. Also, since v(0|r=^o "^ ^QQ^O < 0, it follows that for t > to, v((f)(t)) < v(x(to)) = v(jco) ^ 0. Since by assumption all the eigenvalues of A have negative real parts, we know that the equilibrium JC = 0 of (L) is uniformly asymptotically stable. Thus, lim^-^oo v((/)(0) = 0, which leads to a contradiction. Thus, P must be positive definite. 2. If at least one of the eigenvalues of A has positive real part and no real part of any eigenvalue of A is zero and if (7.15) is satisfied, and if C in (7.7) is negative definite, then P cannot be positive definite; otherwise we could apply Theorem 7.2 to come up with a contradiction. If in particular the real parts of all eigenvalues of A are positive, then P must be negative definite. [Note that in this case the equilibrium JC = 0 of (L) is completely unstable.] Now suppose that at least one of the eigenvalues of A has positive real part, and suppose that any one of the two conditions or both conditions given in (7.15) are not satisfied. Then we cannot construct v(x) given in (7.5) in the manner described above (i.e., we cannot determine P in the manner described above). In this case we form a matrix Ai = A - 81, where / denotes the ^ X n identity matrix and 8 is chosen so that A\ has as many eigenvalues with positive real part as A, but none of the conditions in (7.15) are violated. Then the equation A\P + PAi = C, with C negative definite, can be solved for P, and P is then clearly not positive definite. The derivative of the function v(x) = x^Px is a quadratic form whose matrix
476
is of the form
Linear Systems
(Ai + 81/P
+ P(Ai ^81)
= C + 28P,
which is negative definite for a sufficiently small 8. The function v(x) constructed in this way now satisfies the hypotheses of Theorem 7.3 for system x = Ax. Summarizing the above discussion, we have the following result. THEOREM 7.4. If all the eigenvalues of the matrix A have negative real parts, or if at least one eigenvalue has a positive real part, then there exists a Lyapunov function of the form v(x) = x^Px, P = P^, whose derivative along the solutions of (L) is definite (i.e., it is either negative definite or positive definite). • EXAMPLE 7.4. We consider the system (L) with 0
11
-1 oj' The eigenvalues of A are Ai, A2 = ±j and therefore condition (7.15) is violated. According to Lemma 7.1, the Lyapunov matrix equation A^P + PA = C does not possess a unique solution for a given C. We now verify this for two specific cases. (i) When C = 0, we obtain "0 1
-1] Pn + Pn OJ [pn P22. .Pn
P12I Pii]
[-1
0.
=
-2p\2 .Pn - P22
0 0
Pn - P22 2/712
0 0. '
OT pi2 = 0 and pu = ^22- Therefore, for any a E: R, the matrix P = al is d. solution of the Lyapunov matrix equation. In other words, for C = 0, the Lyapunov matrix equation has in this example denumerably many solutions, (ii) When C = -21, we obtain -2/712 Pn - P22
Pn - P22 2/712
-2 0
0 -2
or Pn = P22 and pu = 1 and pu = - 1 , which is impossible. Therefore, for C -2/, the Lyapunov matrix equation has in this example no solutions at all. It turns out that if all the eigenvalues of matrix A have negative real parts, then we can compute P in (7.7) explicitly. THEOREM 7.5. If all eigenvalues of a real n X n matrix A have negative real parts, then for each matrix C E /?"><", the unique solution of (7.7) is given by /•CO
(7.16) Jo
Proof, If all eigenvalues of A have negative real parts, then (7.15) is satisfied and therefore (7.7) has a unique solution for every C E R^^f^, To verify that (7.16) is indeed this solution, we first note that the right-hand side of (7.16) is well defined, since all eigenvalues of A have negative real parts. Substituting the right-hand side of (7.16) for P into (7.7), we obtain
r°° A^P + PA=\
T f^ T A^e"^ \-C)e'^'dt+\ e^'i-Qe'^'Adt
Jo
All JJ.
Jo
CHAPTER 6
0 at
= e^X-Oe""' which proves the theorem.
= C, 10
6.8 LINEARIZATION In this section we consider nonlinear, finite-dimensional, continuous-time dynamical systems described by equations of the form w = f(w\
(A)
where / G C^(R^, R^). We assume that w = 0 is an equilibrium of (A). In accordance with Subsection 1.11 A, we linearize system (A) about the origin to obtain x = Ax + F(x),
(8.1)
X G R^, where F G C(R^, R^) and where A denotes the Jacobian of f(w) evaluated at w = 0, given by
and where
A = ^(0), dw F(x) = o(\\x\\) as
(8.2) ||;c|| ^ 0.
(8.3)
Associated with (8.1) is the linearization of (A), given by y = Ay,
(L)
In the following, we use the results of Section 6.7 to establish criteria that allow us to deduce the stability properties of the equilibrium w = 0 of the nonlinear system (A) from the stability properties of the equilibrium y = Oof the linear system (L). THEOREM 8.1. Let A G R''^'' be a Hurwitz matrix, let F G C(/?", R""), and assume that (8.3) holds. Then the equilibrium x = 0 of (8.1) [and hence, of (A)] is exponentially stable. Proof. Theorem 7.4 applies to (L) since all the eigenvalues of A have negative real parts. In view of that theorem (and the comments following Lemma 7.1), there exists a symmetric, real, positive definite nX n matrix P such that PA + A^P = C,
(8.4)
where C is negative definite. Consider the Lyapunov function v{x) = x^Px.
(8.5)
The derivative of v with respect to t along the solutions of (8.1) is given by v(x) = x^Px + x^Px = {Ax + F{x)fPx + x^P{Ax + F{x)) = x^Cx + 2x^PF(x).
(8.6)
478 Linear Systems
Now choose 7 < 0 such that x^Cx < 3 / \\x\\^ for all x G R^. Since it is assumed that (^•^) holds, there is a 5 > 0 such that if ||x|| < 5, then ||PF(x)|| < —7 ||x|| for all X e B{d) = {xeR'^ :\\x\\< d}. Therefore, for all x G B{d) we obtain, in view of (8.6), the estimate v ( x ) < 3 r | | x | | 2 - 2 7 | | x | | 2 = 7||x||2. (8.7) Now, let a = minn^^-n^^ v(x). Then a > 0 (since P is positive definite). Take A G (0, a ) , and let C^={xe
B{5) = {xeR'':
\\x\\ < 5} : v(x) < A}.
(8.8)
Then C^ C B{5). [This can be shown by contradiction. Suppose that C^ is not entirely inside B{5). Then there is a point xeC^ that lies on the boundary of 5 ( 5 ) . At this point, v(x) > a > A. We have thus arrived at a contradiction.] The set C^ has the property that any solution of (8.1) starting inC^ ait = to will stay in Q for ^H ^ > ^0 ^ 0- To see this, we let (l){t,to,xo) = 0(r) and we recall that v(x) < 7||x|p,7< 0,x G B{d) D Q - Then v(0(O) < 0 implies that v{^{t)) < v(xo) < A for all ^ > ^o > 0. Therefore, ^{t) G Q for all t>to>0. We now proceed in a similar manner as in the proof of Theorem 7.1 to complete this proof. In doing so, we first obtain the estimate
v((|)(0)<(^^)v((|)(0),
(8.9)
where 7 is given in (8.7) and C2 is determined by the relation ci\\xf
< v{x)=x^Px
(8.10)
Following now in an identical manner as was done in the proof of Theorem 7.1, we have m)\\<(-]
||xo|k^/'(^/^^)(^-^°\
t>to>0,
(8.11)
whenever XQ G 5 ( / ) , where / has been chosen sufficiently small so that B(r') C C^. This proves that the equilibrium x = 0 of (8.1) is exponentially stable. • It is important to recognize that Theorem 8.1 is a local result that yields sufficient conditions for the exponential stability of the equilibrium x = 0 of (8.1); it does not yield conditions for exponential stability in the large. The proof of Theorem 8.1, however, enables us to determine an estimate of the domain of attraction of the equilibrium x = 0 of (A), involving the following steps: 1. 2. 3.
Determine an equilibrium, x^, of (A) and transform (A) to a new system that translates Xe to the origin x = 0 (refer to Section 6.3). Linearize (A) about the origin and determine F(x),A, and the eigenvalues of A. If all eigenvalues of A have negative real parts, choose a negative definite matrix C and solve the Lyapunov matrix equation C=A^P^PA.
4.
Determine the Lyapunov function v{x)
5.
=x^Px.
Compute the derivative of V along the solutions of (8.1), given by v{x)=x^Cx^2x^PF{x).
6.
Determine 5 > 0 such that v(jc) < 0 for all x G B{d) - {0}.
7. Determine the largest A = AM such that Cx^ ^ ^ ( S ) , where CA
479 CHAPTER b
= {x G B!" : v(x) < A}.
Stability
8. Cxyi is a subset of the domain of attraction of the equiUbrium x = 0 of (8.1), and hence, of (A). The above procedure may be repeated for different choices of matrix C given in step (3), resulting in different matrices P/, which in turn may result in different estimates for the domain of attraction, C\ , / G A, where A is an index set. The union of the sets C\^ = Di, D = U/D/, is also a subset of the domain of attraction of the equilibrium x = 0 of (A). THEOREM 8.2. Assume that A is a real nX n matrix that has at least one eigenvalue with positive real part. Let F E C(/?", /?''), and assume that (8.3) holds. Then the equilibrium X = 0 of (8.1) [and hence, of (A)] is unstable. Proof, We use Theorem 7.4 to choose a real, symmetric n X n matrix P such that the matrix PA + A^P = C is negative definite. The matrix P is not positive definite, or even positive semidefinite (refer to the comments following Lemma 7.1). Hence, the function v(x) = x^ Px is negative at some points arbitrarily close to the origin. The derivative of v(x) with respect to t along the solutions of (8.1) is given by (8.6). As in the proof of Theorem 8.1, we can choose a y < 0 such that x^Cx < 3711x11^ for all x E P", and in view of (8.3) we can choose a 5 > 0 such that ||PF(x)|| < y\\x\\ for all x E B(8). Therefore, for all x E B(8), we obtain that
Hx) ^ 3r||x|p - 2r||x|p = rll^lp. Now let G = {xE B(8) : v(x) < 0}. The boundary of G is made up of points where either v(x) = 0 or where ||x|| = 8. Note in particular that the equilibrium x = 0 of (8.1) is in the boundary of G. Now following an identical procedure as in the proof of Theorem 7.3, we show that any solution (/)(0 of (8.1) with 4>(to) = XQ E: G must escape G in finite time through the surface determined by Ikll = 8. Since the above argument holds for arbitrarily small 5 > 0, it follows that the origin x = 0 of (8.1) is unstable. • Before concluding this section, we consider a few specific cases. EXAMPLES.1. The Lienard Equation is given by w + g{w)w + w = 0,
(8.12)
where g E C^(R, R) with ^(0) > 0. Letting xi = w and X2 = w, we obtain Xi = X2
(8.13)
X2 = - X i -
Let x^
(xi, X2), f{xf
g(Xi)X2. g{Xi)X2.
= (/i(x), /2(x)), and let
"1^(0) 1^(0)" 7(0) = A =
dxi
dX2
1^(0) 1^(0) _dX\
=
0 -1
1
•
-^(0).
dX2
Then X = A^ : + [/(x) - Ax] = Ax + F(x\
480 Linear Systems
where
0 l[^(0) - g(Xi)]X2\
F(x)
The origin x = 0 is clearly an equilibrium of (8.12) and hence of (8.13). The eigenvalues of A are given by
JgW
Ai, A2 =
and therefore, A is a Hurwitz matrix. Also, (8.3) holds. Therefore, all the conditions of Theorem 8.1 are satisfied. We conclude that the equilibrium x = 0 of (8.13) is exponentially stable. • EXAMPLE 8.2. We consider the system given by X\ = -Xi
+ X\{x\ + x\)
(8.14)
X2 = -X2 + X2(xj + x\).
The origin is clearly an equilibrium of (8.14). Also, the system is already in the form (8.1) with -1 0
0 -1
F{x) =
\x\{x\ + X2)l
[x2{x\ + xl)y
and condition (8.3) is clearly satisfied. The eigenvalues of A are Ai = - 1 , A2 = - 1 . Therefore, all conditions of Theorem 8.1 are satisfied and we conclude that the equilibrium x^ = (xi, X2) = 0 is exponentially stable; however, we cannot conclude that this equilibrium is exponentially stable in the large. Accordingly, we seek to determine an estimate for the domain of attraction of this equilibrium. We choose C = -I (where / G R^^^ denotes the identity matrix) and we solve the matrix equation A^P + PA = C to obtain P = (1/2)/, and therefore, V(xi, X2) = x^Px = \(x] + xl). Along the solutions of (8.14) we obtain v(xi, X2) = x^Cx +
2x^PF(x)
= -(x^ + xj) + (xi +
xjf.
Clearly, v(xi, X2) < 0 when (xi, X2) ^ (0, 0) and Xj + ^2 < 1. In the language of the proof of Theorem 8.1, we can therefore choose 8 = 1. Now let Ci/2 = {xeR^:
v(xi, X2) = \(x1 + x\) < i}.
Then clearly. Cm C B(8X 8 = 1, in fact Cm = B(8). Therefore, the set {x 1 R^:xj + X2 < 1} is a subset of the domain of attraction of the equilibrium (xi, X2)^ •- 0 of system(8.14). • EXAMPLES.3. The differential equation governing the motion of a pendulum is given by (8.15)
6 -\- asinO = 0, where a > 0 is a constant (refer to Chapter 1). Letting 6 the system description
xi and 6 = X2, we obtain
Xi = X2
X2 =
(8.16) -asinxi.
The points x^.^^ = (0, 0)^ and xf^ = (77, 0)^ are equilibria of (8.16).
(i) Linearizing (8.16) about the equilibrium x^^\ we put (8.16) into the form (8.1) with
481 CHAPTER 6
Stability
0 1 A= -a 0 The eigenvalues of A are Ai, A2 = ±7 V^. Therefore, the results of this section (Theorem 8.1 and 8.2) are not applicable in the present case. (ii) In (8.16), we let y\ ~ x\ - TT and 3^2 = ^2- Then (8.16) assumes the form yi -(2sin(yi + 77).
(8.17)
The point (3^1, 3^2)^ = (0,0)^ is clearly an equilibrium of system (8.17). Linearizing about this equilibrium, we put (8.17) into the form (8.1), where 0 0 1 F(yh yi) = A= -a(sin(3;i +77) + yO. a 0 The eigenvalues of A are Ai, A2 = a, -a. All conditions of Theorem 8.2 are satisfied and we conclude that the equilibrium xf^ = (TT, 0)^ of system (8.16) is unstable. •
PART 2 INPUT-OUTPUT STABILITY OF CONTINUOUS-TIME SYSTEMS 6.9 INPUT-OUTPUT STABILITY We now turn our attention to systems described by the state equations X = A(t)x + B(t)u y = C(t)x + D(t)u,
(9.1)
where A G C(R, T^'^^'^), B G C(R, 7^"^^), C G C(K R^'"''), and D G C(R, RP"""^) [resp., A G C{R^, i?"^^'^), B G C{R^, R'"'"'^), C G C{R+, RP""""), and D G C(R^, RP^^)]. In the preceding sections of this chapter we investigated the internal stability properties of system (9.1) by studying the Lyapunov stability of the trivial solution of the associated system w = A(t)w.
(LH)
In this approach, system inputs and system outputs played no role. To account for these, we now consider the external stability properties of system (9.1), called input-output stability: every bounded input of a system should produce a bounded output. More specifically, in the present context, we say that system (9.1) is boundedinput/bounded-output (BIBO) stable if for all to and zero initial conditions at f = to, every bounded input defined on [to, ^) gives rise to a bounded response on [to, 0°). A bounded matrix D(t) does not affect the BIBO stability of (9.1), while an unbounded D(t) will give rise to an unbounded response to an appropriate constant input. Accordingly, we will consider without any loss of generality the case where D(t) = 0, i.e., throughout this section we will concern ourselves with systems described by equations of the form
482 Linear Systems
X = A{t)x 4- B{t)u (9 2) y = C(t)x. We will find it useful to use a more restrictive concept of input-output stability in establishing various results: we will say that the system (9.2) is uniformly BIBO stable if there exists a constant k> Q that is independent of ^ , such that for all ^ the conditions X{tQ) = 0
||w(0|| ^ 1 ,
t^
to,
imply that ||j(Oli — k for all t > to. (The symbol || • || denotes the Euclidean norm.) It turns out that for the class of problems considered herein, BIBO stability and uniform BIBO stability amount to the same concepts. (We will not, however, prove this assertion here.) Accordingly, we will phrase all subsequent results in terms of uniform BIBO stability, rather than BIBO stability. These results will involve the impulse response matrix of (9.2) given by
Har) = f^«*(''-)^(->' ^;^' [ U,
(9.3)
t < T,
and the controllability and observability Gramian given, respectively, by W(to, ^i) -
0 ( % t)B(t)B^(t)^^(to,
t)dt
(9.4)
J/o ch
and
M(% ^i) =
^{t, tof C{tf C{t)^{t, to) dt,
(9.5)
In these results we will establish sufficient conditions for uniform BIBO stability of (9.2) and also necessary and sufficient conditions for uniform BIBO stability of (9.2). Furthermore, we will present results that make a connection between the uniform BIBO stability of (9.2) and the Lyapunov exponential stability of the equilibrium w = 0 of (LH). In view of the latter results, we will usually assume that to ^ 0. At the end of this section we will also present specialized stability results for the time-invariant systems described by equations of the form X = Ax + Bu
(9.6)
y ^ Cx,
where A G Z^'^^^", B E Z^^^^'^, and C G RP''''. Associated with system (9.6) is the free system described by equations of the form P = Ap,
(L)
Recall that for system (9.6) the impulse response matrix is given by H(t) = Ce'^'B,
t > 0,
= 0,
/ < 0,
(9.7)
and the transfer function matrix is given by H(s) = C(sl -A)-^B. (9.8) THEOREM 9.1. The system (9.2) is uniformly BIBO stable if and only if there exists a finite constant L > 0 such that for all t and to, with t > to,
\\H{t,T)\\dr£L.
(9.9)
483 CHAPTER 6 :
Stability
The first part of the proof of Theorem 9.1 (sufficiency) is straightforward. Indeed, if ||M(r)|| < 1 for all t^tQ and if (9.9) is true, then we have for all f ^ fo that
lb(Oll =
Hit, T)U{T) dr to
<
\\H(t,T)u(T)\\dr Jto rt \\H(t,T)\\\\u(T)\\dT I
< \
\\H(t,T)\\dT^L.
Jto
Therefore, system (9.2) is uniformly BIBO stable. In proving the second part of Theorem 9.1 (necessity), we simplify matters by first considering in (9.2) the single-variable case (n = 1) with the input-output description given by y(t) =
(9.10)
h(t,T)u(r)dT.
For purposes of contradiction, we assume that the system is BIBO stable, but no finite L exists such that (9.9) is satisfied. Another way of stating this is that for every finite L, there exist to = to(L) and ti = ^i(L), ti > to, such that \h(tuT)\dr>
L.
to
We now choose in particular the input given by r+1 if/z(^,T)>0, u(t) = \ 0 ifh(t,r) = 0, [-1 ifh(t,T)<0,
(9.11)
to ^ t ^ ti. Clearly, \u(t)\ < 1 for all t ^ to. The output of the system att = ti due to the above input, however, is rti y(ti)
=
rti h{ti, T)U{T)dr
Jto
=
\h{ti,T)\dT
> L,
Jto
which contradicts the assumption that the system is BIBO stable. The above can now be extended to the multivariable case. In doing so, we apply the single-variable result to every possible pair of input and output vector components, we make use of the fact that the sum of a finite number of bounded sums will be bounded, and we recall that a vector is bounded if and only if each of its components is bounded. We leave the details to the reader. In the preceding argument we made the tacit assumption that u is continuous, or piecewise continuous. However, our particular choice of u may involve nondenumerably many switchings (discontinuities) over a given finite-time interval. In such cases, u is no longer piecewise continuous; however, it is measurable (in the
484 Linear Systems
Lebesgue sense). This generalization can be handled, though in a broader mathemati^al setting that we do not wish to pursue here. The interested reader may want to refer, e.g., to the books by Desoer and Vidyasagar [5], Michel and Miller [17], and Vidyasagar [25] and the papers by Sandberg [21] to [23] and Zames [26], [27] for further details. From Theorem 9.1 and from (9.7) it follows readily that a necessary and sufficient condition for the uniform BIBO stability of system (9.6) is the condition \\H(t)\\dt<^. (9.12) Jo COROLLARY 9.1. Assumc that the equilibrium w = 0 of (L^) is exponentially stable and suppose there exist constants ^ > 0 and y > 0 such that for all t, \\B{t)\\ < j8 and ||C(0|| ^ 7- Then system (9.2) is uniformly BIBO stable. Proof, Under the hypotheses of the corollary, we have
^0
H{t,T)dTl < [ \\H(t,T)\\dT II -'^0 = f' \\C(t)^(t, T)B(T)\\dT < 7i8 f' |cD(^, T)|| Jr.
Since the equilibrium w = 0 of (LH) is exponentially stable, there exist 6 > 0, A > 0 such that ||0(r, T)|| < 8e~^^^~^\ t > r. Therefore, \\H{t,T)\\dT^ [
ypde-^^'-'-Ur
JtQ
to
ySp A A for all T, t with r > r. It now follows from Theorem 9.1 that system (9.2) is uniformly BIBO stable. • As indicated earlier, we seek to establish a connection between the uniform BIBO stability of (9.2) and the exponential stability of the trivial solution of (LH). We will accomplish this by means of an intermediate result for systems described by equations of the form X = A(t)x + B(t)u ^' ^'' y = X.
(9.13)
Before stating and proving the next result, we recall that if 5 G R^^^^ J E R^^^ are symmetric, then the notation S > 0 signifies that S is positive definite, and the notation S > T indicates that the matrix S - T is positive definite, i.e., S - T > 0. Also, if Q(t) = Q(tf G C(R, /^"><"), the condition that there is a constant r] > 0 such that Q(t) > 17/ for all ^ E i? is equivalent to the statement that z^Q(t)z ^ r]\\z\\^ for all f E /? and all z E R"". Next, suppose that there is a constant a > 0 such that ||A(0|| ^ oc for all t ^ R, and let $(^, r) denote the state transition matrix of (LH). In the proof of the next result we will require the estimate ||(I)(r,T)||< ^"^
k - T | < 8.
(9.14)
To obtain this estimate we let (f)(t, r, ^) = 0 ( 0 denote the solution of (LH) with (/)(T) == ^, and we compute
^Mfm
= ^U(tt = ktfm
at
+ itfm
——^
at
CHAPTER 6 :
=
^{t) + (i){tf Ait)
Stability
for all t > T. Letting ||(^(0|F = v{t), we have
— < lav,
V(T) = ll^lp,
which yields
or
mt
^ e-^'-^m ^
e-'\U
which in turn yields (9.14). The case when r > n s treated similarly. THEOREM 9.2. Suppose that there exist positive constants a, /3, e, and 8 such that for all t, \\A{t)\\ < a, \\B{t)\\ < j8, and ^(^0, ^o, +5) ^ e/, where / denotes the n X w identity matrix and W(-) denotes the controllability Gramian given in (9.4). Then the system (9.13) is uniformly BIBO stable if and only if the trivial solution of {LH) is exponentially stable. Proof, Under the above hypotheses, it follows from Corollary 9.1 that if the trivial solution of {LH) is exponentially stable, then system (9.13) is uniformly BIBO stable. Conversely, assume that the system (9.13) is uniformly BIBO stable and assume that the hypotheses of the theorem are satisfied with the given constants a, /3, 6, and e. The assumption W(^, t^ + 8)> el ensures that W~^{tQ, to + 8) exists, is bounded, and is independent of ^o- We now consider / = [
[(I>(T, r])B(r])Bir]Y^(T, r]f d7]W\T
- 8, r).
(9.15)
Jr-8
Since B(r]) is bounded and since ^ ( r , r]) is bounded over |T - r/l < 5, there exists a constant c > 0 such that ||5(r/)^cI)(T, r])'^W~\T - 8, T)|| < c.
(9.16)
Premultiplying (9.15) by ^(t, r) and using the bound (9.16) and the properties of norms, we obtain ||cDaT)||
m,rj)B(rj)\\dri.
(9.17)
JT-8
Since system (9.13) is uniformly BIBO stable, there exists a /: > 0 such that mt,r])B(v)\\dri
(9.18)
Jt-n8
for all positive integers n, where k is independent of n and t. From (9.18) it follows that m.V)B(v)\\dv t-nd
= f
mt,v)B(ri)\\d7]+[
Jt-S
\\^(t,7])B(rj)\\drj Jt-28
rt-n8+8
+ •••+
\\^(t,r))B{'n)\\dy)
(9.19)
Jt-n8
From (9.17) to (9.19) it follows that c-^W^it, Oil + c-^W^iU t - 8)\\ + •" + c-^W^it, t-n8
+ 5)11 \\^{t,y))B{y])\\d7)
486
Since the above is true for any positive integer n, we have
Linear Systems
c'Wl^it, or
t)\\ + \\^(t, t - d)\\ + • • • + \\^{t, t - nd + 8)\\ + • • • ) < ^
\\^{t, Oil + l l ^ a t - 5)11 + • • • + \\^{t, t-nd
+ 5)11 + "•
To complete the proof, we must show that \[^ \\^{t, y]% drj is finite. We will accomplish this by showing that |_^g ||^(^, 17)11 drj is finite for any positive integer. To this end we observe that for any n, t given, there exists a positive integer m such that ||a>(/, T/)|Mr7 < f -n8
\m,7j)\\drj.
Jt—m8
If we apply the Mean Value Theorem for Integrals to f
||cDar;)||Jr^= f
Jt-m8
mt,v)\\d7i+C
Jt~8
110^77)11^7, Jt-28
rt-(m-l)8
+ •••+
mt,v)\\dv> Jt-m8
we obtain for fjt ^ [t - id, t - (i - 1)5], / = 1 , . . . , m, that ||(Da 77)11^7, < [ -nS
110^77)11^7,
Jt-m8
< 5[||oa 77011 + l|oa 772)11 + • • • + iioa 77^)111 Now invoking Theorem 5.5(iv) of this chapter, we conclude that the equilibrium w = 0 of (LH) is exponentially stable. • THEOREM 9.3. Suppose that there exist positive constants a, jS, and y such that ||A(0|| ^ «, II^Wll — P^ and ||C(0|| ^ y for all t and assume that there exist positive constants e 1, 62, 5i, and 52 such that for all to, W(to, ^o + ^i) — ^ i^ and M(to, to + 82) ^ 62/, where M(') denotes the observability Gramian given in (9.5). Then the system (9.2) is uniformly BIBO stable if and only if the equilibrium w = 0 of (LH) is exponentially stable. Proof, Uniform exponential stability of the trivial solution of (LH) and the hypotheses of this theorem imply the uniform BIBO stabiHty of system (9.2) by Corollary 9.1. To complete the proof, we show that the hypotheses of the theorem and the BIBO stability of system (9.2) imply the exponential stability of the equilibrium w = 0 of (LH). To set up a contradiction, assume that uniform BIBO stability of system (9.2) does not imply exponential stability of the trivial solution of (LH). Then by Theorem 9.2, it must not imply uniform BIBO stability of system (9.13). For if there is no bound on the state, then there can be no bound on the output. To see this, let w = 0 on ^ < r < r + 5 and obtain t+8
rt+8
|b(T)|pdr = xitf[\
r)dr]x{t)
Jt
= x(tfM(t,
t + 8)x(t) <
max d\\y(T)f, t
and therefore, max ||>;(T)f > d-'\\x(t)f\,^AM(t,
t + 5)],
t
where Amin[^(^, t + 5)1 denotes the smallest eigenvalue of M(t, t + 5). Thus, if the state X is not bounded for all bounded w, then y will also not be bounded. This shows that the
uniform BIBO stability of (9.2) implies the uniform BIBO stability of (9.13). Applying Theorem 9.2, we conclude that the equilibrium w = Oof (LH) is exponentially stable. • EXAMPLE 9.1. We consider the system described by the scalar equations 1
. X+ U
^^^'
(9.20)
y = X.
This system is clearly controllable and observable. The zero-input response of this system is determined by the differential equation • ^
Jw,
w(to) = xo,
to > 0.
(9.21)
It is easily verified that the solution of (9.21) is given by ct>(t,to,xo)= \ ^ ^ o
(9.22)
[refer to Eqs. (4.5) and (4.6)]. The origin w = 0 is the only equilibrium of (9.21), and as seen from (9.22), this equilibrium is uniformly stable and asymptotically stable; however, it is not uniformly asymptotically stable, and hence, it is not exponentially stable. The state transition matrix of system (9.21) is given by 1'(Uo) =
y ^ .
With ^ = 0 and bounded input u(t) = 1, / > 0, the zero-state response of system (9.20) is r ^ 1 _(_ /T-
r t
y(t,to>xo)=
Jo
^(t,T)u(T)dT
= ^
=
T-r-dr Jo ^ + ^
.
(9.23)
Summarizing, even though the zero input response of system (9.20) tends to zero as ^ -^ 00, the hypotheses of Theorem 9.3 are not satisfied since this decay is not exponential and uniform [i.e., the origin w = 0 of (9.21) is not exponentially stable]. In accordance with Theorem 9.3, we cannot expect the zero-state response of system (9.18) to be bounded for arbitrary bounded inputs. This is evident from expression (9.23). • Next, we consider in particular the time-invariant system (9.6). For this system it is easily verified that Theorem 9.3 reduces to the following appealing result that connects the uniform BIBO stability of system (9.6) and the exponential stability of the trivial solution of (L). THEOREM 9.4. Assume that the time-invariant system (9.6) is controllable and observable. Then system (9.6) is uniformly BIBO stable if and only if the trivial solution of (L) is exponentially stable. • EXAMPLE9.2. Consider the system X = Ax + Bu
where
0 "^ " .1i
1 o0.r
y = Cx, 0 B ^ "= .1. i r'
^ = ^^
-H-
The eigenvalues of A are A i = 1, A2 = - 1 , and therefore the equilibrium w of the system w = Aw is unstable. The state transition matrix of this system is
487 CHAPTER 6* Stability
488 Linear Systems
0(^, 0) =
and the impulse response of the system is H{t) = COaO)5 =
-e-\
Thus, even though the equihbrium w = 0 of w = Aw is unstable, the system is uniformly BIBO stable. The reason for this is that the system is not observable, as is verified by noting that C 1 -1 CA -1 1 which is singular. Thus, the hypotheses of Theorem 9.4 are not valid, with the consequence that the unstable mode of the system is not observable at the system output. • EXAMPLE9.3. We no w consider the system X = Ax ^- Bu
y = Cx, where A and B are as in Example 9.2 and C = [\ 2]. The eigenvalues and the state transition matrix of this system are identical to those of the system given in Example 9.2. This system is both controllable and observable, and the impulse response is H{t) =\e'-
\e-\
Since the system is controllable and observable and since the equilibrium w = 0 of the system w = Aw is unstable, it follows from Theorem 9.4 that the system cannot be uniformly BIBO stable. This can be verified directly by inspecting H(t) above. • Next, we recall that a complex number Sp is a pole of H(s) = [hij(s)] if for some pair (/, j), we have \hij(sp)\ = oo (refer to the definition of pole in Section 3.5). If each entry of H(s) has only poles with negative real values, then, as shown in Chapter 2, each entry of H(t) = [hij(t)] has a sum of exponentials with exponents with real part negative. It follows that the integral \\H(t)\\dt 0
is finite, and any realization of H(s) will result in a system that is uniformly BIBO stable. Now conversely, if r 00
\\H(t)\\dt 0
is finite, then the exponential terms in any entry of H(t) must have negative real parts. But then every entry of H(s) has poles whose real parts are negative. We have proved the following result. THEOREM 9.5. The time invariant system (9.6) is uniformly BIBO stable if and only if all poles of the transfer function H(s) given in (9.8) have only poles with negative real parts. •
PART 3 STABILITY OF DISCRETE-TIME SYSTEMS ^•H^^H^^H
489 ^^^^6: Stability
6.10 DISCRETE-TIME SYSTEMS In this section we address the Lyapunov stability of an equilibrium of discrete-time systems (internal stability) and the input-output stability of discrete-time systems (external stability). We could establish results for discrete-time systems that are analogous to practically all the stability results that we presented for continuous-time systems. Rather than follow such a plan, we will instead first develop Lyapunov stability results for nonlinear discrete-time systems and then apply these to obtain results for Hnear systems. This will broaden the reader's horizon by providing a glimpse into the qualitative analysis of dynamical systems described by nonlinear ordinary difference equations. In the Exercise section we ask the reader to imitate these results in establishing Lyapunov stability results for dynamical systems described by nonlinear ordinary differential equations. To keep our presentation simple and manageable, we will confine ourselves throughout this section to time-invariant systems. Among other issues, this approach will allow us to avoid most of the issues involving uniformity. This section is organized into seven subsections. In the first subsection we provide essential preliminary material. In the second we establish results for the stability, instability, asymptotic stability, and global asymptotic stability of an equilibrium and boundedness of solutions of systems described by autonomous ordinary difference equations. These results are utilized in the third and fifth subsections to arrive at stability results of an equilibrium for linear time-invariant systems described by ordinary difference equations. In the fourth subsection we briefly address a result for discrete-time systems, called the Schur-Cohn criterion, which is in the same spirit as the Routh-Hurwitz criterion is for continuous-time systems. The results of the second and fifth subsections are used to develop Lyapunov stability results for Hnearizations of nonlinear systems described by ordinary difference equations in the sixth subsection. In the last subsection we present results for the input-output stability of time-invariant discrete-time systems.
A. Preliminaries We concern ourselves here with finite-dimensional discrete-time systems described by difference equations of the form x(k -\-l) = Ax(k) + Bu(k)
(10.1)
y(k) = Cx{k), where A G /^'^X", B G W''''^, C G /?^><^ k ^ k^, and k, ko^Z+. Since (10.1) is time-invariant, we will assume without loss of generahty that ko = 0, and thus, X : Z+ -> 7?^ J : Z+ -> RP, and w : Z+ ^ R"^. The internal dynamics of (10.1) under conditions of no input are described by equations of the form
490
x(k + 1) = Axik). ^
(10.2)
Such equations may arise in the modehng process, or they may be the consequence of the hnearization of nonlinear systems described by equations of the form x(k + 1) = g(x(k)),
(10.3)
where g : R^ -> R^. For example, if ^ E C^{R^, R^), then in linearizing (10.3) about, e.g., x = 0, we obtain x(k + 1) = Ax(k) + f(x(k)\
(10.4)
where A = (df/dx)(x)\^^Q and where f : R^ ^ R^ is o(\\x\\) as a norm of x (e.g., the Euclidean norm) approaches zero. Recall that this means that given e > 0, there is a S > 0 such that \\f(x)\\ < e||x|| for all ||x|| < 8. As in Section 6.9, we will study the external qualitative properties of system (10.1) by means of the BIBO stability of such systems. Since we are dealing with time-invariant systems, we will not have to address any issues of uniformity. Consistent with the definition of input-output stability of continuous-time systems, we will say that the system (10.1) is BIBO stable if there exists a constant L > 0 such that the conditions jc(0) -
0
\\u{k)\\ < 1,
yt > 0,
imply that \\y{k)\\ < L for all fc > 0. We will study the internal qualitative properties of system (10.1) by studying the Lyapunov stability properties of an equilibrium of (10.2). We will accomplish this in a more general context by studying the stability properties of an equilibrium of system (10.3). Since system (10.3) is time-invariant, we will assume without loss of generality that ko = 0. As in Chapters 1 and 2, we will denote for a given set of initial data x(0) = XQ the solution of (10.3) by (p^k, XQ). When XQ is understood or of no importance, we will frequently write (/)(fc) in place of (/)(A:, XQ). Recall that for system (10.3) [as well as systems (10.1), (10.2), and (10.4)], there are no particular difficulties concerning the existence and uniqueness of solutions, and furthermore, as long as g in (10.3) is continuous, the solutions will be continuous with respect to initial data. Recall also that in contrast to systems described by ordinary differential equations, the solutions of systems described by ordinary difference equations [such as (10.3)] exist in general only in the forward direction of time (k > 0). We say that Xe E R^ is an equilibrium of system (10.3) if cl)(k, Xe) = Xe for all /: > 0, or equivalently, g{Xe) = Xe.
(10.5)
As in the continuous-time case, we will assume without loss of generality that the equilibrium of interest will be the origin, i.e., Xe = 0. If this is not the case, then we can always transform (similarly as in the continuous-time case) system (10.3) into a system of equations that have an equilibrium at the origin. Also, as in the case of continuous-time systems, we will generally assume that the equilibrium of (10.3) under study is an isolated equilibrium.
EXAMPLE 10.1. The system described by the equation x{k + 1) = x{k)[x{k)
491
- 1]
has two equihbria, one dX Xe\ = 0 and another at Xg2 = 2.
•
E X A M P L E 1 0 . 2 . The system described by the equations
xi{k + 1) = ^lik) X2(k + 1) = -^i(^) has an equiUbrium at xf = (0, 0).
•
Throughout this section we will assume that the function g in (10.3) is continuous, or if required, continuously differentiable. The various definitions of Lyapunov stability of the equilibrium x == 0 of system (10.3) are essentially identical to the corresponding definitions of Lyapunov stability of an equilibrium of continuous-time systems described by ordinary differential equations, replacing ^ G /?+ by ^ G Z+. Since system (10.3) is time-invariant, we will not have to explicitly address the issue of uniformity in these definitions. We will concern ourselves with stability, instability, asymptotic stability, and asymptotic stability in the large of the equilibrium X = Oof (10.3). We say that the equilibrium x == 0 of (10.3) is stable if for every e > 0 there exists 3. 8 = S(e) > 0 such that \\4>(k, xo)\\ < e for all /: ^ 0 whenever ||xo|| < 8. If the equilibrium x = 0 of (10.3) is not stable, it is said to be unstable. We say that the equilibrium x = 0 of (10.3) is asymptotically stable if (i) it is stable, and (ii) there exists an 17 > 0 such that if ||xo|| < 17, then lim^^^oo ||)(^, xo)|| = 0. If the equilibrium x = 0 satisfies property (ii), it is said to be attractive, and we call the set of all XQ G R^ for which x = 0 is attractive the domain of attraction of this equilibrium. If x = 0 is asymptotically stable and if its domain of attraction is all of R^, then it is said to be asymptotically stable in the large or globally asymptotically stable. Finally, we say that a solution of (10.3) through XQ is bounded provided there is a constant M such that ||(/)(/:, xo)|| ^ M for all ^ > 0. In establishing results for the various stability concepts enumerated above, we will make use of auxiliary functions v G C{R^, R), called Lyapunov functions. We define the first forward difference ofv along the solutions (9/(10.3) as Dv{x) = v(g(x)) - v(x).
(10.6)
To see that this definition makes sense, note that along any solution (/)(^) of (10.3) we have v[ct,(k + 1)] - v[cl>(k)] = v[g{cl>(k))] - vmk)]
^^Q ^^
= Dvmk)] for all fc > 0. Note that in evaluating the first forward difference of v along the solutions of (10.3), we need not know explicitly the solution (/>(/:) of (10.3). We will require several characterizations of the Lyapunov functions. We say that a function v G C(R^, R) is positive definite if v(0) = 0 and if v(x) > 0 for all x G B{j]) - {0} for some ry > 0. [Recall that B{j]) = {x G /?" : ||x|| < ry}.] The function V is negative definite if - v is positive definite. A function v G C{R^, R) is said to be positive semidefinite if v(0) = 0 and if v(x) > 0 for all x G ^(17), and it is said to be negative semidefinite if - v is positive semidefinite. A positive definite function v is said to be radially unbounded if v(x) > 0 for all x G 7?" - {0} and if limiij^iuoo v(x) =
CHAPTER 6: Stabihty
492 Linear Systems
00. Finally, a function v E C(R^, R) is said to be indefinite if v(0) = 0 and if in every neighborhood of the origin v assumes positive and negative values. B. Lyapunov Stability of an Equilibrium In establishing various Lyapunov stability results of the equilibrium x = 0 of system (10.3), we will find it useful to employ a preliminary result that is important in its own right. To present this result, we require some additional concepts given in the following. We say that a subset A C 7^" is positively invariant [with respect to system (10.3)] if g{A) C A, where g{A) = {y E. R"" \ y = g{x) for some x G A}. Thus, if X EL A, then g{x) G A. EXAMPLE 10.3. It is clear that any set consisting of an equilibrium solution of (10.3) is positively invariant. Also, the set {^{k, XQ), k E Z^} [where 4> is the solution of (10.3), with (/)(/co) = ^o], which is called i\\Qpositive orbit of XQ for (10.3), is positively invariant. In particular, for arbitrary initial conditions x(0)^ == (^i(O), ^2(0)), the system given in Example 10.2 has iho periodic solution with period 4 and positive orbit given by the set A = {(xi(0), X2(0))^, (X2(0), -xmV,
(-xm,
-^2(0))^, (-X2(0),
and in general we have (f)(k, XQ) = (t)(k -\- 3, xo\ k = 0, 1, 2, 3, positively invariant.
xmYl
The set A is clearly •
Next, consider a specific solution (/)(/:, XQ) for system (10.3). A point y E R"^ is called Si positive limit point of 4>{k, XQ) if there is a subsequence {hi} of the sequence {k}, /: > 0, such that (f){ki, XQ) -^ y. The set of all positive limit points of (f){k, XQ) is called the positive limit set (o(xo) of (/)(/:, XQ), or simply the co-limit set. Before stating and proving our first result, which is a preliminary result, we recall that a sequence {x^} C R^ is said to approach a set A C R^, if d(x^, A) = inf {||x^ - y\\ : y E A} approaches zero asfc^ oo [d(x, y) denotes the distance function defined in Subsection I.IOC]. THEOREM 10.1. If the solution (t)(k, XQ) of (10.3) is bounded for all k E Z+, then (o(xo) is a nonempty, compact, positively invariant set. Furthermore, (t)(k, XQ) -^ w(jco) as /: ^
00.
Proof, The complement of (O(XQ) is open, and therefore, the set (w(xo) is closed. Since 4>(k, XQ) is bounded, ||(^(/:, xo)|| ^ M for some fixed M. Therefore, (o(xo) is bounded and IIJII < M for all y E (O(XQ). Thus, (JO(XO) is compact. Since (f>(k, XQ) is bounded, the Bolzano-Weierstrass Theorem guarantees the existence of at least one limit point, and therefore, (x)(xo) is not empty (refer to Subsection 1.5A). Next, let y E co(xo) so that (/)(A:/, XQ) -^ y as / ^ OO. Since the function g in (10.3) is continuous, it follows that (l)(ki + I, XQ) = g{4^{ki, XQ)) -^ g{y) or g{y) E O){XQ). The set O){XQ) is positively invariant [with respect to system (10.3)]. Also, d{(j){k, XQ), CL>) is bounded since both (j){k, XQ) and w(xo) are bounded. To set up a contradiction, assume that d{(f){k, XQ), O){XQ)) does not converge to zero. Then there is a subsequence [ki] such that (/)(/:/, XQ) -> y and d((p(ki, XQ), (i^(xo)) -> a > 0. But then y E ct)(xo), d{(j){ku XQ), (I>{XQ)) < d{(f)(ku XQ), y) -^ 0, and therefore, d(
THEOREM 10.2. The equilibrium x = 0 of system (10.3) is stable if there exists a positive definite function v such that Dv is negative semidefinite.
493 CHAPTER 6-
Proof, We take r/ > 0 so small that v(x) > 0 for all x E 5(7]) - {0} and Dv(x) < 0 for all X E 5(77) [recall that Birf) = {x ^ R^ \ ||x|| < r/}]. Let e > 0 be given. There is no loss of generality in choosing 0 < e < 17. Let m = min {v(x) : \\x\\ = e}. Then m is positive since we are taking the minimum of a positive continuous function over a compact set. Let G = {x B R^ : v(x) < mil}, which may consist of several disjoint connected sets called connected components of G. Let Go denote the connected component of G that contains the origin x = 0. Then both G and Go are open sets. Now if XQ E GO, then DV{XQ) < 0 and so v(^(xo)) ^ V{XQ) < m/2, and therefore, g^xo) E G. Since XQ and the origin x = 0 are both in the same component of G, then so are ^(0) = 0 and g(xo). Therefore, Go is an open positively invariant set containing x = 0 and contained in B{e). Since v is continuous there is a 5 > 0 such that B{b) C Go. Therefore, if xo E B{d), then Xo E Go and ^(xo) E Go C B{e). This shows that the equilibrium x = 0 of system (10.3) is stable. •
Stability
EXAMPLE 10.4. For the system given in Example 10.2 we choose the function v(xi, X2) = x\ + x\. Then Dv(xi, X2) = v(g(xi, X2)) - v(xi, X2) = (xif
+ (-xif
- xj-
xl = 0.
Therefore, by Theorem 10.2 the equiUbrium x = 0 of the system is stable.
•
By using a similar argument as in the proof of Theorem 10.2 we can prove the boundedness result given below. We will not present the details of the proof. THEOREMIO.3. Ifvis radially unbounded and Dv(x) < 0 on the set where ||x|| > M (M is some constant), then all solutions of system (10.3) are bounded. • EXAMPLE 10.5. For the system given in Example 10.2 we choose the radially unbounded function v(xi, X2) = x^ + X2. As shown in Example 10.4, Dv{x\, X2) ^ 0 for all X E /?^. It follows from Theorem 10.3 that all the solutions of the system are bounded. • By definition, the equilibrium x = 0 of system (10.3) is asymptotically stable if it is stable and attractive. Theorem 10.2 provides a set of sufficient conditions for the stability of the trivial solution of system (10.3). In the next result, known as LaSalle 's Theorem or the Invariance Principle, we present a method that enables us to determine the attractivity of a set. If this set consists only of the equilibrium x: = 0, this result yields a method of determining the attractivity of the trivial solution x = Q of system (10.3). Let V E C{R'', R) and let v(x) - c. In the following, we let v~^{c) = {x B R"" : v(x) = c}. THEOREM 10.4. Let V E C(/?", R) and let G C R"". Assume that (i) Dv(x) < 0 for all X G G, and (ii) the solution 4>(k, XQ) of (10.3) is in G for all A: > 0 and is bounded for all ^ > 0. Then there is a number c such that (/)(^, xo) ^ M Pi v~^(c), where M is the largest positively invariant set contained in the set E = {x G R'^ : Dv(x) = 0} D G. Proof. Since (f)(k, XQ) is bounded and in G, it follows that co(xo) 7^ 0 , CL)(XO) C G , and (/)(/:, xo) tends to a>(xo). Now v{(j){k, XQ)) is nonincreasing with increasing k and bounded from below. Therefore, v(4>(k, XQ)) -> c. If y G CL)(XO), there is subsequence {ki} of the sequence {k} such that (/>(/:/, xo) -^ y and therefore v{(f)(ki, xo)) -> v{y) or v(y) = c. Therefore, V{(X){XQ)) = c [i.e., v(y) = c for all y E a;(xo)] or ^(xo) C v~^(c). Since v(6o(xo)) = c and a;(xo) is positively invariant, it follows that Dv(o;(xo)) = 0 [i.e..
494 Linear Systems
Dv(y) = 0 for all y G co(xo)]. Therefore, 0(^, XQ) -^ (x)(xo) C {x E /?" : Dv(x) = 0} Pi G fl v^^c). Since co(xo) is positively invariant, it now follows that a)(xo) CM. • The next results, which are direct consequences of Theorems 10.2, 10.3, and 10.4, were originally established by Lyapunov. COROLLARY 10.1. The equilibrium x = 0 of (10.3) is asymptotically stable if there exists a positive definite function v such that Dv is negative definite. Proof. Since Dv is negative definite and v is positive definite, it follows from Theorem 10.2 that the equilibrium x = 0 is stable. From the proof of that theorem there is an arbitrarily small neighborhood Go of the origin that is positively invariant. We can make Go so small that v(x) > 0 and v(x) < 0 for all x G Go - {0}. So, given any xo G Go it follows from the invariance principle (Theorem 10.4) that (pik, XQ) tends to the largest invariant set in Go H {Dv(x) = 0} = {x = 0} since Dv is negative definite. Therefore, the equilibrium x = 0 is asymptotically stable. • COROLLARY 10.2. The equilibrium x == 0 of (10.3) is asymptotically stable in the large (or globally asymptotically stable) if there exists a positive definite function v that is radially unbounded and if Dv is negative definite on R"", i.e., Z)v(O) = 0 and Z)v(x) < 0 for all X 7^ 0. • Proof, From Theorem 10.3 all solutions of system (10.3) are bounded. The proof now follows by modifying the proof of Corollary 10.1 in the obvious way. • EXAMPLE 10.6. Consider the system described by the equations xi(k+
1) =
X2(k + 1) =
ax2(k) 1 + xi(y^)2
bxi(k) 1 + X2(k)^'
where it is assumed that a^ < I and b^ < I. We choose a Lyapunov function v{x\, X2) = x\ + x\ that is positive definite and radially unbounded. Along the solutions of this system we compute the first forward difference as
^ v t e , . . ) = ^ ^ + ^ ^ - ( x ? + 4) 2\2
(1 + x\y
1
Xo +
b^
1 2\2
(1 + 4>
< {a^ -\)xl + (b^ - l)x?. Since by assumption a^ < I and Z?^ < 1, it follows that Dv(xi, X2) < 0 for all x^ = (xi, X2) 7^ 0 and Dv(xi, X2) = 0 when x = 0. It follows from Theorem 10.3 that all solutions of the system are bounded, and from Corollary 10.2 it follows that the equilibrium X = 0 of the system is globally asymptotically stable. • EXAMPLE 10.7. We reconsider the system given in Example 10.6 under the assumption that a^ ^ I and Z?^ < 1, but a^ + b'^ 7^ 2. Without loss of generality we consider the case a^ < 1 and b^ = 1. As in Example 10.6, we again choose v(xi, X2) = x^ + x^, and from the computations in that example we see that Dv(xu X2) < (a^ - \)x\ + Q? - \)x\ = {a^ - \)x\ < 0, (xi, X2)^ G R^. It still follows from Theorem 10.3 that all solutions of the system are bounded.
Since Dv(xi, X2) is not negative definite, but negative semidefinite, we cannot apply Corollary 10.2 to establish the asymptotic stability of the equilibrium x = 0. So let us try to use Theorems 10.2 and 10.4 to accomplish this. From the former we conclude that the equilibrium is stable. Using the notation of Theorem 10.4 we note that E = {(xi, 0)^}, which is the xi-axis. Now g((xi, OY) = (0, bxiY = (0, xi)^ [where g(') is defined in (10.3)], and therefore, the only invariant subset of (E) is the set consisting of the origin. All conditions of Theorem 10.4 are satisfied (with G = R^), and we conclude that the equilibrium x = 0 of the system is globally asymptotically stable. • THEOREM10.5. Let Dv be positive definite and assume that in every neighborhood of the origin there is x such that v(x) > 0. Then the equilibrium x = 0 of system (10.3) is unstable. [Alternatively, let Dv be negative definite and assume that in every neighborhood of the origin there is x such that v(x) < 0. Then the equilibrium x = 0 of system (10.3) is unstable.] Proof, To set up a contradiction, assume that x = 0 is stable. Choose e > 0 so small so that Dv{x) > 0 for all x G B{e) - {0}, and choose 6 > 0 so small so that if XQ G B(8) then 4>{k, XQ) G B{e) for all ^ > 0. By hypothesis there is a point XQ G B{3) such that v(xo) > 0- Since (j){k, XQ) is bounded and remains in 5(e), the solution (f){k, XQ) will tend to its Hmit set {x G /?" : Dv{x) = 0} Pi B{e) = {0}. Since (/)(/:, XQ) -^ 0, we have v((l)(k XQ)) -» v(0) = 0. But Dv((l)(k, xo)) > 0 and therefore v((/)(^, XQ)) > 0, and thus v(4>(k, Xo)) ^ v(cl)(k-l, Xo)) > ••• > v(xo) > 0. We have thus arrived at a contradiction that proves the theorem. • EXAMPLE 10.8. The system described by the equation x(k +1) = 2x(k) has an equilibrium x = 0. The function v(x) = x^ is positive definite, and along the solutions of this system we have Z)v(x) = 4x^-x^ = 3x^, which is also positive definite. The conditions of Theorem 10.5 are satisfied and the equilibrium x = 0 of the system is unstable. •
C. Linear Systems In proving some of the results of this section, we require a result for system (10.2) that is analogous to Theorem 3.1 of Chapter 2. As in the proof of that theorem, we note that the linear combination of solutions of system (10.2) is also a solution of system (10.2), and hence, the set of solutions {4>: Z^ XR^ -^ R^} constitutes a vector space (over F = Ror F = C). The dimension of this vector space is n. To show this, we choose a set of linearly independent vectors XQ, ..., XQ in the n-dimensional x-space (R^ or C^) and we show, in an identical manner as in the proof of Theorem 3.1 of Chapter 2, that the set of solutions (/)(fc, XQ), i = 1 , . . . , n, is linearly independent and spans the set of solutions of system (10.2). (We ask the reader in the Exercise section to provide the details of the proof of the above assertions.) This yields the following result. THEOREM 10.6. The set of solutions of system (10.2) over the time interval Z+ forms an ^-dimensional vector space. • Incidentally, if in particular we choose (/)(^, e^), i = I,.. .,n, where e\ i = 1 , . . . , n, denotes the natural basis for R^, and if we let ^{k, yto = 0) = 0(A;) = [ct){k, e^),..., (pik, e")], then it is easily verified that the n X n matrix 0(A:) satisfies
495 CHAPTER 6*
stabilitv
496
the matrix equation
Linear Systems
ci>(k + 1) - A^(k\ and that (5(fc) = A^,k^
c|)(0) = /
0 [i.e., 3>(/:) is the state transition matrix for system (10.2)].
THEOREM 10.7. The equiHbrium x = 0 of system (10.2) is stable if and only if the solutions of (10.2) are bounded. Proof. Assume that the equilibrium x = 0 of (10.2) is stable. Then for 6 = 1 there is a 8 >0 such that \\cl)(k, xo)\\ < 1 for all ^ > 0 and all ||jco|| < 6. In this case U(k, xo)\\ = WA'xoW =
A^XQS
m\\
M 8
< 8
for all xo 7^ 0 and all k > 0. Using the definition of matrix norm [refer to (10.17) of Chapter 1] it follows that ||A^|| < 8~\ A: > 0. We have proved that if the equilibrium X = 0 of (10.2) is stable, then the solutions of (10.2) are bounded. Conversely, suppose that all solutions (j6(/:, XQ) = A'^XQ are bounded. Let{^\ . . . , ^"} denote the natural basis for ^-space and let \\(p(k, e^)\\ < jBj for all k > 0. Then for any vector XQ = S " = i cty^^ we have that
ll^(^,xo)||
^^aj(f)(k,
e^)
Xk-|i8;^(max^,.)Xl^il 7=1
7= 1
^
7=1
0,
c\\m
for some constant c. For given 6 > 0, we choose 8 = elc. Then, if ||xo|| < 8, we have \^{k, xo)|| < c||xo|| < 6 for all /: > 0. We have proved that if the solutions of (10.2) are bounded, then the equilibrium x = 0 of (10.2) is stable. • THEOREM 10.8. The following statements are equivalent: (i) The equilibrium x = 0 of (10.2) is asymptotically stable. (ii) The equilibrium x = 0 of (10.2) is asymptotically stable in the large, (iii) lim^^oo ||Ai = 0. Froof Assume that statement (i) is true. Then there is anry > 0 such that when ||xo|| ^ 17, then ^{k, XQ) -> 0 as /: -^ 00. But then we have for any XQ ^ 0 that (^{k, xo) = A^xo
17x0
Ikoll
Ikoll
0 as /: ^ 00.
It follows that statement (ii) is true. Next, assume that statement (ii) is true. Then for any e > 0 there must exist a K = K{e) such that for ?^k> K we have that \^{k, xo)|| = ||A^xo|| < 6. To see this, let {^^ . . . , ^"} be the natural basis for R^. Thus, for a fixed constant c > 0, if XQ = (a\,..., anf andif ||xo|| ^ l,thenxo 2:;.i«y^^'and2:;=iK ji -^ c. For eachy there is a Kj - Kj{e) such that \A^e^ <6/cfory^> TTy. Define TT K(€) = max {Kj(e) : j = 1 , . . . , n}. For ||xo| < 1 and k > K we have that
ll^'^oll = 7= 1
7= 1
By the definition of matrix norm [see (10.17) of Chapter 1], this means that ||A^|| < e for k> K. Therefore, statement (iii) is true. Finally, assume that statement (iii) is true. Then ||A^|| is bounded for all /c > 0. By Theorem 10.7, the equilibrium x = 0 is stable. To prove asymptotic stability, fix e > 0.
If llxoll < 77 = 1, then \\(t)(k, xo)\\ < \\A^\\ \\xo\\ ^ 0 as y^ ^ 00. Therefore, statement (i) is true. This completes the proof. • To arrive at the next result, we make reference to the results of Subsection 2.7E. Specifically, by inspecting the expressions for the modes of system (10.2) given in (7.50) and (7.51) of Chapter 2, or by utilizing the Jordan canonical form of A [refer to (7.54) and (7.55) of Chapter 2], the following result is evident. THEOREM 10.9. (i) The equilibrium x = 0 of system (10.2) is asymptotically stable if and only if all eigenvalues of A are within the unit circle of the complex plane (i.e., if A i , . . . , A„ denote the eigenvalues of A, then \Xj\ < I, j = 1 , . . . , ^). In this case we say that the matrix A is Schur stable, or simply, the matrix A is stable. (ii) The equilibrium x = 0 of system (10.2) is stable if and only if \Xj\ < 1, j = 1,.. .,n, and for each eigenvalue with |Aj| = 1 having multiplicity rij > 1, it is true that
}^^S^^^'-'^^"'^''-^^"^\='^
h...,nj-
1.
(iii) The equilibrium x = 0 of system (10.2) is unstable if and only if the conditions in (ii) above are not true. • Alternatively, it is evident that the equilibrium x = 0 of system (10.2) is stable if and only if all eigenvalues of A are within or on the unit circle of the complex plane, and every eigenvalue that is on the unit circle has an associated Jordan block of order 1. EXAMPLE 10.9. (i) For the system in Example 10.2 we have A -
0 -1
11 0
The eigenvalues of A are Ai, A2 = ± v - 1 - According to Theorem 10.9, the equilibrium x = 0 of the system is stable, and according to Theorem 10.7 the matrix A^ is bounded for all k^O. (ii) For system (10.2) let 0 -1 The eigenvalues of A are Ai, A2 = ±1/V2. According to Theorem 10.9, the equilibrium X = 0 of the system is asymptotically stable, and according to Theorem 10.8, lim^^^ooA^ = 0. (iii) For system (10.2) let A The eigenvalues of A are Ai, A2 = ± V3/2. According to Theorem 10.9, the equilibrium X = 0 of the system is unstable, and according to Theorem 10.7, the matrix A^ is not bounded with increasing k. (iv) For system (10.2) let A =
1 0
1 1
The matrix A is a Jordan block of order 2 for the eigenvalue A = 1. Accordingly, the equilibrium x = 0 of the system is unstable (refer to the remark following Theorem 10.9) and the matrix A^ is unbounded with increasing k. •
497 CHAPTER 6:
Stability
498
D. The Schur-Cohn Criterion
Linear Systems
In this section we present a method, called the Schur-Cohn criterion, that enables us to determine whether or not the roots of a polynomial with real coefficients given by + • • • + ^0,
an > 0,
(10.8)
lie inside of the unit circle in the complex plane by examining the polynomial coefficients, rather than solving for the roots. This method provides us with an efficient means of studying the stability of the equilibrium x = 0 of system (10.2) by applying it to the characteristic polynomial of the matrix A. The Schur-Cohn criterion is in the same spirit [for discrete-time systems (10.2)] as the Routh-Hurwitz criterion [for continuous-time systems (L)]. In presenting this criterion, we will require the concept of inners of a square matrix A, defined as the matrix itself, and all the matrices obtained by omitting successively the first and last rows and the first and last columns. In the following, we depict the inners for a matrix A E R^^^ and a matrix A G i?6x^: an
an
cin
«21 1 Cl22 « 2 3
A =
<2i4 <324
^25
^ 3 1 1 ^32 1 (233
(234
(241 1 ^ 4 2
(243
CI44
(245
«52
«53
^54
ass.
Cl5l
<335
A =
ai3
ai4
(215
(216
ail
^23
(224
^25
^26
<332
«33
^42
<^43
^35 (245
<^36
(241
^34 (244
asi
asi
<353
asA
ciss
(256
(262
^63
(264
aes
^66
(211
an
<321 (231
(246
L
(261
A matrix is said to be positive innerwise if the determinants of all of its inners are positive. THEOREM 10.10. (SCHUR-COHN CRITERION) A neccssary and sufficient condition that the polynomial (10.8) with real coefficients has all its roots inside the unit circle in the complex plane is that (i) p(l) > 0 and ( - 1)V(-1) > 0, (ii) the following (n — I) X (n — 1) matrices are both positive innerwise: 0 0
0
(3o
(3o
(31
0
Cln-l (31
0
as (32
(10.9)
as
Cln~\
0
Cln.
(30
(3^n-l
ai
^n-1
We will not present a proof of Theorem 10.10. The interested reader should consult Jury [91 cited in the reference section for a proof of this result. EXAMPLE 10.10. For the 2 X 2 matrix A=
(311
an
CI2I
ail.
the characteristic polynomial is given by p(\)
= A^ + (3iA + (3o.
We have p(l) = I + ai + ao, p(-l) = 1 - ai + ao, and Af = 1 ± (^o > 0. Therefore, 499 the roots of p(X) he within the unit circle of the complex plane if and only if |ao| < 1 CHAPTER 6: and |ai| < 1 + ao. • Stability EXAMPLElO.il. For the 3 X 3 matrix ai3
an
an
dii
«22
<223
.asi
^32
<^33
the characteristic polynomial is given by p(X) = A^ + (22 A^ + ai\ + ao. We have p(l) = 1 + ^2 + ai + <3o, -p(-l) det{^^) =
= I - a2 + ai ao, and
1 -ao > 0 . a2 ± ao 1 ±ai
Therefore, the roots of p(X) lie within the unit circle of the complex plane if and only if Wo + a2\ < I + ai and \ai - aoa2\ < I - al. • E. The Matrix Lyapunov Equation In this subsection we apply the Lyapunov theorems of Subsection B to obtain another characterization of stable matrices. This gives rise to the Lyapunov matrix equation. Returning to system (10.2) we choose as a Lyapunov function v(x) = x^Bx,
B = B^,
(10.10)
and we evaluate the first forward difference of v along the solutions of (10.2) as v{x{k + 1)) - v{x{k)) = x(k + lfBx(k
+ 1) -
= x{kfA^BAx{k) = x{kf{A^BA
-
x(kfBx(k) x{kfBx{k)
B)x(k\
and therefore, Dv(x) = x^{A^BA - B)x = x^Cx, where
A^BA -B
= Q
C^ = C.
(10.11)
Invocation of Theorem 10.2, Corollary 10.2, and Theorem 10.5, readily leads to the following results. THEOREM 10.11. (i) The equilibrium x = 0 of system (10.2) is stable if there exists a real, symmetric, and positive definite matrix B such that the matrix C given in (10.11) is negative semidefinite. (ii) The equilibrium x = 0 of system (10.2) is asymptotically stable in the large if there exists a real, symmetric, and positive definite matrix B such that the matrix C given in (10.11) is negative definite. (iii) The equilibrium ;\f = 0 of system (10.2) is unstable if there exists a real, symmetric matrix B that is either negative definite or indefinite such that the matrix C given in (10.11) is negative definite. • In applying Theorem 10.11, we start by choosing (guessing) a matrix B having certain desired properties and we then solve for the matrix C, using equation (10.11).
500 Linear Systems
If C possesses certain desired properties (i.e., it is negative definite) we can draw appropriate conclusions by applying one of the results given in Theorem 10.11; if not, we need to choose another matrix B. This approach is not very satisfactory, and in the following we will derive results that will allow us (as in the case of continuous-time systems) to construct Lyapunov functions of the form v{x) = x^Bx in a systematic manner. In doing so, we first choose a matrix C in (10.11) that is either negative definite or positive definite, and then we solve (10.11) for 5. Conclusions are then made by applying Theorem 10.11. In applying this construction procedure, we need to know conditions under which (10.11) possesses a (unique) solution B for any definite (i.e., positive or negative definite) matrix C. We will address this issue next. We first show that if A is stable, i.e., if all eigenvalues of matrix A [in system (10.2)] are inside the unit circle of the complex plane, then we can compute B in (10.11) explicitly. To show this, we assume that in (10.11) C is a given matrix and that A is stable. Then (A^)^+i5A^+i - {A^fBA^
=
(A^fCA^
and summing from ^ = 0 to / yields / A^BA -B
+ {A^fBA^
- A^BA +
B =
^{A^fCA^ k=0 I
or
(A^y^^BA^^^
-B==
^(A^)^CA^. k=0
Letting / ^ oo^ we obtain B = -^(A^)^CA\
(10.12)
k=0
It is easy to verify that (10.12) is a solution of (10.11). We have - A ^ ^(A^)^CA^ k=0
or
-A^CA + C - (A^fCA^
+ A^CA - (A^fCA^
A + ^(A^)^CA^
- C
k=0
+ (A^fCA^
= C.
Therefore (10.12) is a solution of (10.11). Furthermore, if C is negative definite, then B is positive definite. Combining the above with Theorem lO.ll(ii) we have the following result: THEOREM 10.12. If there is a positive definite and symmetric matrix B and a negative definite and symmetric matrix C satisfying (10.11), then the matrix A is stable. Conversely, if A is stable, then, given any symmetric matrix C, Eq. (10.11) has a unique solution, and if C is negative definite then B is positive definite. • Next, we determine conditions under which the system of equations (10.11) has a (unique) solution B = B^ E R^^^ for a given matrix C = C^ E: R^^^, To accomplish this, we consider the more general equation A 1 X A 2 - X = C, where A\ E R"^^^^ A^ G R""^"", and X and C are m X n matrices.
(10.13)
LEMMA 10.1. Let Ai G /^'^^^^ and A2 G /?"><^ ThenEq. (10.13) has a unique solution X E R^^^ for a given C E /^'^x" if and only if no eigenvalue of A\ is a reciprocal of an eigenvalue of A2. Proof, We need to show that the condition on Ai and A2 is equivalent to the condition that A1XA2 = X implies X ^ 0. Once we have proved that A1XA2 = X has the unique solution X = 0, then it can be shown that (10.13) has a unique solution for every C, since (10.13) is a linear equation. Assumefirstthat the condition on Ai and A2 is satisfied. Now A1XA2 = X implies that A\~^XA\~^ = X and A{X
= A\XA\~^
for y^ > 7 > 0.
Now for a polynomial of degree k, k
7=0
we define the polynomial of degree k,
/(A) = X^i^'"'' = ^'/^fT\ 7=0 from which it follows that p{A,)X =
A\Xp\A2l
Now let (/)/(A) be the characteristic polynomial of A/, / = 1, 2. Since (f)\ (A) and (/>2(A) are relatively prime, there are polynomials p{\) and ^(A) such that p{\)cj>,{K) + q{k)cf>l{K) = 1. Now define (j>{X) = q(X)(f)l(X) and note that (/)*(A) = ^*(A)(/)2(A). It follows that (/)*(A2) = 0 and (p(Ai) = I. From this it follows that A1XA2 = X impUes X = 0. To prove the converse, we assume that A is an eigenvalue of Ai and A~^ is an eigenvalue of A2 (and hence, is also an eigenvalue of A^). Let Aix^ = Xx^ and A^x^ = X'^x^, x^ ¥^ 0, and x^ 7^ 0. Define X = {x\x\ xlx\ ..., JC^X^). Then X 7^ 0 and A1XA2 = X. • To construct v{x) by using Lemma 10.1, we must still check the definiteness of B. To accomplish this, we utilize Theorem 10.11. 1. If all eigenvalues of A [for system (10.2)] are inside the unit circle of the complex plane, then no reciprocal of an eigenvalue of A is an eigenvalue, and Lemma 10.1 gives another way of showing that Eq. (10.11) has a unique solution B for each C if A is stable. If C is negative definite, then B is positive definite. This can be shown as was done for the case of linear ordinary differential equations. 2. If at least one of the eigenvalues of A is outside the unit circle of the complex plane and if no reciprocal of an eigenvalue of A is an eigenvalue, and if C in (10.11) is negative definite, then B cannot be positive definite; otherwise we could apply Theorem lO.ll(iii) to come up with a contradiction. If in particular, all eigenvalues of A are outside the unit circle of the complex plane, then B must be negative definite. [In this case the equilibrium of (10.2) is completely unstable.} Now suppose that at least one of the eigenvalues of A is outside of the unit circle in the complex plane and suppose that the conditions of Lemma 10.1 are violated (i.e., an eigenvalue of A is the reciprocal of an eigenvalue of A). Then we cannot
501 CHAPTER 6*
Stability'
502
Linear Systems
construct v(x) given in (10.10) in the manner described above (i.e., we cannot determine B in the manner described above). To overcome this difficulty, we form in this case the matrix
Ao = (1 +
(10.14)
lir"'A,
and we choose |j8| arbitrarily small and in such a manner so that AQ has no eigenvalues on the unit circle and has the same number of eigenvalues outside of the unit circle of the complex plane as matrix A. Then AQ satisfies the conditions of Lemma 10.1. For every given C = C^ ^ R^^^ we can now solve the equation AlBAo
(10.15)
B = C
to obtain a unique matrix B. We use this matrix to form the Lyapunov function (10.10). Now since A = (1 + /3)^^^Ao, the first forward difference Dv(x) of the Lyapunov function v(x) = x^Bx along the solutions of (10.2) yields Dv{x)
= X'^IAIBAQ = x^{C
- B +
+ I3AIBAO)X
[5AIBAO]X =
(10.16)
x^Cx,
where C is given in (10.15). It now follows that if C is negative definite (positive definite), then we can choose |jS| sufficiently small so that C will also be negative definite (positive definite). Summarizing the above discussion, we have proved the following result. THEOREM 10.13. If all the eigenvalues of the matrix A are within the unit circle of the complex plane, or if at least one eigenvalue is outside the unit circle of the complex plane, then there exists a Lyapunov function of the form v{x) = x^Bx, B = B^, whose first forward difference along the solutions of system (10.2) is definite (i.e., it is either negative definite or positive definite). • We conclude this subsection with some specific examples. EXAMPLE 10.12. (i) For system (10.2) let A=
0 n •1 0
Let B = I, which is positive definite. From (10.11) we obtain C = A^A - I
ro
1
-11 r 0 1 o| [-1 0
1 0 0 1
0 0 0 0
It follows from Theorem lO.ll(i) that the equilibrium x = 0 of this system is stable. This is the same conclusion that was made in Example 10.9. (ii) For system (10.2) let A= Choose
which is positive definite. From (10.11) we obtain A^BA - B = L
0 -11 r^ 01 r 0 3 ^ 1 oj L0^ ^3 J [-1 2
n 2
oJ
r^ 0" 3 ^ 0 ^ L^
3-
-1 [ 0
0" -ij
which is negative definite. It follows from Theorem 10.1 l(ii) that the equilibrium x = 0 of this system is asymptotically stable in the large. This is the same conclusion that was made in Example 10.9(ii). (iii) For system (10.2) let 0
A=
-i 0
-3
Choose -1 0
C =
0 -1
which is negative definite. From (10.11) we obtain 1 2
L
m22-bn) [ \bn
-3^ pll bn'] [ 0 0^ [b\2 Z722J L-3
0
C = A^BA -B =
bn bu
0.
bn b22.
0]
-1 0
\bn {\bn - b22).
n
2
-ij '
which yields -8 0
B=
0 -1
which is also negative definite. It follows from Theorem 10.11 (iii) that the equilibrium X = 0 of this system is unstable. This conclusion is consistent with the conclusion made in Example 10.9(iii). (iv) For system (10.2) let
I 1
0 3 The eigenvalues of A are Ai = | and A2 = 3. According to Lemma 10.1, for a given C, Eq. (10.13) does not have a unique solution in this case since Ai = I/A2. For purposes of illustration we choose C = - / . Then -I = A'^BA - B 1^11 ^bn
1^11 bn + €>bi2 + 8^22.
.1
[^11
bn]
3j [bl2
^22]
H 1^
[0 3. =
bn bn
bn b22.
-1 0 0 -- 1 .
which shows that for C = - / , Eq. (10.13) does not have any solution (for B) at all.
R Linearization In this subsection we determine conditions under which the stability properties of the equilibrium w = 0 of the linear system w(k+ 1) = Aw(k)
(10.17)
determine the stability properties of the equilibrium x = 0 of the nonlinear system x(k + 1) = Ax(k) + f(x(k)) under the assumption that f(x)
(10.18)
= o(\\x\\) as ||x|| -^ 0 (i.e., given e > 0 there exists
503 CHAPTER 6 :
Stability
504 Linear Systems
S > 0 such that \\f(x(k))\\ < e\\x(k)\\ for a l U > 0 and all \\x(k)\\ < 8). [Refer to the discussion concerning Eqs. (10.2) to (10.4) in Subsection A of this section.] THEOREM 10.14. Assumethat/ G C(R'', R"") and thai f(x) is o(\\x\\)sis\\x\\ -» 0. (i)If A is stable (i.e., all the eigenvalues of A are within the unit circle of the complex plane), then the equilibrium x = 0 of system (10.18) is asymptotically stable, (ii) If at least one eigenvalue of A is outside the unit circle of the complex plane, then the equilibrium X = 0 of system (10.18) is unstable. Proof, (i) Assume that A is stable. Then for any negative definite matrix C, the equation A^BA
-B
=C
has a unique positive definite solution B. Let v(x) — x^Bx. Along the solutions of (10.18) we compute the first forward difference Dv(x) as Dv(x) = x^Cx + 2x^A^Bf(x)
+ v(/(x)).
This allows us to estimate [since f(x) is 6>(||x||) as ||x|| -^ 0] Dv(x) < AM(C)|WP + 2||x|| ||A|| \\B\\ \\f(x)\\ + v(/(x)) < AM(C)||X|P + 2||A|| \\B\\e\\xf + AM(5)6^||X|P = [AM(C) + 2||A|| \\B\\e + XM(B)e^\xf
(10.19)
= y||x|p
for all X E B(8) and for some 6 > 0, where AM(C) < 0, XM(B) > 0 denote the largest eigenvalues of C and B, respectively. We can make e as small as desired by choosing 8 sufficiently small, resulting in y < 0. Therefore, Dv(x) is negative definite and the equilibrium x = 0 of system (10.18) is asymptotically stable. (ii) Assume that A has at least one eigenvalue outside the unit circle of the complex plane. Following the procedure given in proving Theorem 10.13 [refer to Eqs. (10.14) to (10.16)] we construct a Lyapunov function v(x) = x^Bx whose first forward difference along the solutions of (10.17) is given by Dv(x) = x^Cx [refer to (10.16)]. We choose C to be negative definite. Then B is indefinite, and in every neighborhood of the origin there are x G R^ such that v(x) = x^Bx < 0. Next, we evaluate the first forward difference of v along the solutions of (10.18) to obtain [identically as in (10.19)], Dv(x) < [AM(C) + 2||A|| \\B\\€ + AM(B)6^]||X|P = y\\xf
for all X G B(8) for some 6 > 0, where C is defined by (10.14) to (10.16). As in the proof of part (i), we can force e to be as small as desired by choosing 8 sufficiently small, resulting again in y < 0. Therefore, Dv(x) is negative definite. It follows from Theorem 10.5 that the equilibrium x = 0 of (10.18) is unstable. • Before concluding this subsection, we consider some specific examples. EXAMPLE 10.13. (i) Consider the system Xi(k+l)
= -{X2(k) + Xi(kf
X2(k + 1) = -xi(k)
+ xi(kf
Using the notation of (10.18) we have A =
0
-1
-1
0
f{M, X2)
+ X2(kf + X2{kf.
The linearization of (10.20) is given by
505
w(k + 1) - Aw(k).
(10.21)
From Example 10.9(ii) [and Example 10.12(ii)] it follows that the equilibrium w = 0 of (10.21) is asymptotically stable. Furthermore, in the present case f(x) = o(\\x\\) as \\x\\ -> 0, Therefore, in view of Theorem 10.14, the equilibrium x = 0 of system (10.20) is asymptotically stable. (ii) Consider the system Xi(/:+ 1) = -^X2(k) + xi(kf + X2(kf X2(k + 1) = -3x1 (y^) + xt(k) - X2{k)\ Using the notation of (10.17) and (10.18), we have in the present case 0
A=
i
X-\
2
3
/(•^b
T" X 9
-^2)
0.
Since A is unstable [refer to Example 10.12(iii) and Example 10.9(iii)] and since f{x) = o(\\x\\) as ||x|| -> 0, it follows from Theorem 10.14 that the equilibrium x = 0 of system (10.22) is unstable. •
G. Input-Output Stability We conclude this chapter by considering the input-output stability of discrete-time systems described by equations of the form x(k + 1) - Ax(k) + Bu(k) (10.23) y(k) - Cx(kl where all matrices and vectors are defined as in (10.1). Throughout this subsection we will assume that ^0 = 0. -^(0) = 0, and ^ > 0. As in the continuous-time case, we say that system (10.23) is BIBO stable if there exists a constant c > 0 such that the conditions x(0) = 0 \\u(k)\\ < 1,
k^O,
imply that ||3;(^)|| < c for all fc > 0. The results that we will present involve the impulse response matrix of (10.23) given by H(k) =
fCA^-^B, [ 0 ,
k>0, (10.24) yk < 0,
and the transfer function matrix given by H(z) = C{zl-
AY^B.
(10.25)
Recall that n
y(n) = ^H(n-
k)u{k).
(10.26)
CHAPTER 6: Stability
506 Linear Systems
Associated with system (10.23) is the free dynamical system described by the equa^^^^ p{k + 1) - Ap{k),
(10.27)
THEOREM 10.15. The system (10.23) is BIBO stable if and only if there exists a constant L > 0 such that for all n > 0,
X 11^(^)11 ^L.
(10.28)
As in the continous-time case, the first part of the proof of Theorem 10.15 (sufficiency) is straightforward. Specifically, if ||w(fe)|| < 1 for all /: > 0 and if (10.28) is true, then we have for all n > 0, \\y{n)\\ = \\^H(n
- k)u(k)\\ ^J^\\H(n
-
kMk)\\
Therefore, system (10.23) is BIBO stable. In proving the second part of Theorem 10.15 (necessity), we simplify matters by first considering in (10.23) the single-variable case (n = 1) with the system description given by t
y(t) = ^h(t-
k)u{k),
t > 0.
(10.29)
k=0
For purposes of contradiction, we assume that the system is BIBO stable, but no finite L exists such that (10.28) is satisfied. Another way of expressing the last assumption is that for any finite L, there exists t = ki(L) = ki such that ^1
^
\h(ki ~ k)\ > L.
k=0
We now choose in particular the input u given by +1 u(k) = ^ [-1
0
if h(t-
k)>0,
ifh(t-k)
= 0,
ifh(t-
k)<0,
0 < k ^ ki. Clearly, \u(k)\ < 1 for all k ^ 0. The output of the system ait =^ ki due to the above input, however, is y(ki) = ^
h(ki - k)u(k) = ^
\h(ki - k)\ > L,
which contradicts the assumption that the system is BIBO stable.
The above can now be extended to the multivariable case. In doing so we apply the single-variable result to every possible pair of input and output vector components, we make use of the fact that the sum of a finite number of bounded sums will be bounded, and we note that a vector is bounded if and only if each of its components is bounded. We leave the details to the reader. Next, we establish a connection between the asymptotic stability of the equilibrium p = 0 of system (10.27) and the BIBO stability of system (10.23). First, we note that the asymptotic stability of the equilibrium x = 0 of system (10.27) implies the BIBO stability of system (10.23) since the sum \k-l\
B
k=l
is finite. The main task in proving the converse to the above statement is to show that controllability and observability of system (10.23) and the finiteness of the sum Er=i IIC'A^"^^!! imply the finiteness of the sum S^Li ||^^~^ ||- ^^ ^^^ ^^ assume that the system (10.23) is BIBO stable. Then, since k-l
< y \\CA''-^J+^^B\\
\\yik)\
— .^
7=0
II
I
we must have that the power series y \\CA^-^B\\ X-u II
II
k=\
is finite (i.e., absolutely convergent), and this implies that limCA^"^5 = 0.
(10.30)
From (10.30) we can conclude that lim CAA^-^B = lim CA^'^AB = lim CA^B = 0, and repeating, we arrive at lim{CA^)A^-\A'B)
=0,
^,r = 0 , 1 , . . . , n - 1.
We can write this as C CA A^-^[5,A5,--,A^-^5]=0.
lim
(10.31)
CAn-\ If we now assume that (10.23) is controllable (from-the-origin) and observable, then we can select n linearly independent columns of the controllability matrix to form an invertible n x n matrix W and n linearly independent rows of the observability matrix to form an invertible n x n matrix M. Using (10.31) we conclude that limMA^-V = 0
(10.32)
507 CHAPTER 6:
Stability
508 Linear Systems
which yields M-'[limMA^-^w]w-^
= limA^-^
= 0.
(10.33)
From (10.33) we can conclude that the equilibrium j:? = 0 of (10.27) is asymptotically stable. To prove this assertion we assume to the contrary that j:? = 0 of (10.27) is not asymptotically stable. This implies that A has an eigenvalue A with |A| > 1. We assume the case when A is real and leave the case when A is complex as an exercise for the reader. Let ry be an eigenvector associated with A (which must be real). Then A^T] = X^rj,k> 0. If 17 = XQ denotes an initial condition, then the corresponding solution of (10.27) is given by x{k) = A^T] = X^TJ, fc > 0, which does not go to zero as ^ —> 00, i.e., limy^_,oo A^~^ 7^ 0, which contradicts (10.33). {Note: The above proof solves Exercise 6.31 for the case where the eigenvalues are real.) We have thus arrived at the following result. THEOREM 10.16. Assume that system (10.23) is controllable and observable. Then system (10.23) is BIBO stable if and only if the equilibrium /? = 0 of system (10.27) is asymptotically stable. • Next, we recall that a complex number Zp is dipole of H(z) = [hjiz)] if for some (/, j) we have \hij(zp)\ = ^ (refer to Section 3.5 for the definition of a pole). If each entry of H(z) has only poles with modulus (magnitude) less than 1, then, as shown in Chapter 2, each entry of H(k) = [hij(k)] consists of a sum of converging terms. It follows that under these conditions the sum 00
k=0
is finite, and any realization of H(z) will result in a system that is BIBO stable. Conversely, if 00
k=0
is finite, then the terms in every entry of H(k) must be convergent. But then every entry of ^(z) has poles whose modulus is within the unit circle of the complex plane. We have proved the final result of this section. THEOREM 10.17. The time-invariant system (10.23) is BIBO stable if and only if the poles of the transfer function H(z) =
C(zI-A)-^B
are within the unit circle of the complex plane.
•
6.11 SUMMARY In this chapter we first addressed the stability of an equilibrium of continuous-time finite-dimensional systems (Part 1). In doing so, we first introduced the concept of equilibrium and defined several types of stability in the sense of Lyapunov (Sections 6.3 and 6.4). Next, we established several stability conditions of an equilibrium for
linear time-varying systems {LH) in terms of the state transition matrix and for linear time-invariant systems L in terms of eigenvalues (Section 6.5). In Section 6.6 we established several geometric and algebraic stability criteria for nth-order, linear, time-invariant systems [including the Leonhard-Mikhailov stability criterion (Theorem 6.1), the gap and position stability criterion (also called the interlacing stability criterion) (Theorem 6.2), and the Routh-Hurwitz criterion (Theorem 6.4)]. Next, we established various stability conditions for linear time-invariant systems that are phrased in terms of the Lyapunov Matrix Equation for system (L) (Section 6.7). In Section 6.8 we established conditions under which the asymptotic stability and the instability of an equilibrium for a nonlinear time-invariant system (A) can be deduced from the linearization of (A). Then in Part 2 we addressed the input-output stability of time-varying and timeinvariant linear, continuous-time, and finite-dimensional systems (Section 6.9). For such systems we established several conditions for bounded input bounded output stability (BIBO stability) and we related some of these to the stabiHty properties of an equilbrium. The chapter concluded with Part 3 (Section 6.10), where we addressed the Lyapunov stability and the input-output stability of (time-invariant) systems. For such systems, we established results that are analogous to the stability results of continuous-time systems considered in Parts 1 and 2. [Among other topics, we discussed in Subsection 6.10D the Schur-Cohn criterion (which is analogous to the Routh-Hurwitz criterion for continuous-time systems), and in Subsection 6.10E we established stability criteria involving the Lyapunov Matrix Equation for the discrete-time case.]
6.12 NOTES The initial contributions to stability theory that took place toward the end of the last century are primarily due to physicists and mathematicians (Lyapunov [14]), while input-output stability is the brainchild of electrical engineers (Sandberg [21] to [23], Zames [26], [27]). Sources with extensive coverage of Lyapunov stability theory include, e.g., Hahn [7], Khalil [11], LaSalle [12], LaSalle andLefschetz [13], Michel and Miller [17], Michel and Wang [18], Miller and Michel [19], and Vidyasagar [25]. Input-output stability is addressed in great detail in Desoer and Vidyasagar [5] and Vidyasagar [25]. For a survey that traces many of the important developments of stability in feedback control, refer to Michel [15]. In the context of linear systems, nice sources on both Lyapunov stability and input-output stability can be found in numerous texts, including Brockett [2], Chen [3], DeCarlo [4], Kailath [10], and Rugh [20]. In developing our presentation, we found the texts by Brockett [2], Hahn [7], LaSalle [12], Miller and Michel [19], and Rugh [20] especially helpful. For a proof of the Schur-Cohn criterion, and other related results, refer to the elegant book by Jury [9]. The background material summarized in the second section is developed in most standard linear algebra texts, including the classic books by Birkhoff and MacLane [1], Gantmacher [6], and Halmos [8]. For more recent references on this subject, refer, e.g., to the books by Michel and Herget [16] and Strang [24].
509 CHAPTER 6:
Stability
510
6.13
Linear Systems
REFERENCES 1. G. Birkhoff and S. MacLane, A Survey of Modern Algebra, Macmillan, New York, 1965. 2. R. W. Brockett, Finite Dimensional Linear Systems, Wiley, New York, 1970. 3. C. T. Chen, Linear System Theory and Design, Holt, Rinehart and Winston, New York, 1984. 4. R. A. DeCarlo, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1989. 5. C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output Properties, Academic Press, New York, 1975. 6. F. R. Gantmacher, Theory of Matrices, Vols. I, II, Chelsea, New York, 1959. 7. W Hahn, Stability of Motion, Springer-Verlag, New York, 1967. 8. P. R. Halmos, Finite Dimensional Vector Spaces, Van Nostrand, Princeton, NJ, 1958. 9. E. I. Jury, Inners and Stability of Dynamical Systems, Robert E. Krieger Publisher, Malabar, PL, 1982. 10. T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980. 11. H. K. KJialil, Nonlinear Systems, Macmillan, New York, 1992. 12. J. P. LaSalle, The Stability and Control of Discrete Processes, Springer-Verlag, New York, 1986. 13. J. P. LaSalle and S. Lefschetz, Stability by Liapunov's Direct Method, Academic Press, New York, 1961. 14. M. A. Liapounoff, "Probleme generate de la stabilite de mouvement," Ann. Fac. Sci. Toulouse, Vol. 9, 1907, pp. 203-474. (Translation of a paper published in Comm. Soc. Math. Kharkow 1893, reprinted in Ann. Math. Studies, Vol. 17, 1949, Princeton, NJ.) 15. A. N. Michel, "Stability: the common thread in the evolution of feedback control," IEEE Control Systems, Vol. 16, 1996, pp. 50-60. 16. A. N. Michel and C. J. Herget, Applied Algebra and Functional Analysis, Dover, New York, 1993. 17. A. N. Michel and R. K. Miller, Qualitative Analysis of Large Scale Dynamical Systems, Academic Press, New York, 1977. 18. A. N. Michel and K. Wang, Qualitative Theory of Dynamical Systems, Marcel Dekker, New York, 1995. 19. R. K. Miller and A. N. Michel, Ordinary Differential Equations, Academic Press, New York, 1982. 20. W. J. Rugh, Linear System Theory, Second Edition, Prentice-Hall, Englewood CHffs, NJ, 1996. 21. I. W. Sandberg, "On the L2-boundedness of solutions of nonlinear functional equations," BellSyst. Tech. J., Vol. 43, 1964, pp. 1581-1599. 22. I. W. Sandberg, "A frequency-domain condition for stability of feedback systems containing a single time-varying nonlinear element," Bell Syst. Tech. J., Vol. 44, 1974, pp.1601-1608. 23. I. W. Sandberg, "Some results on the theory of physical systems governed by nonlinear functional equations," Bell Syst. Tech. J., Vol. 44, 1965, pp. 821-898. 24. G. Strang, Linear Algebra and its Applications, Harcourt, Brace, Jovanovich, San Diego, 1988. 25. M. Vidyasagar, Nonlinear Systems Analysis, 2d edition. Prentice Hall, Englewood Cliffs, NJ, 1993. 26. G. Zames, "On the input-output stability of time-varying nonlinear feedback systems. Part I," IEEE Transactions on Automatic Control, Vol. 11, 1966, pp. 228-238. 27. G. Zames, "On the input-output stability of time-varying nonlinear feedback systems. Part II," IEEE Transactions on Automatic Control, Vol. 11, 1966, pp. 465-476.
6.14 EXERCISES 6.1. Determine the set of equilibrium points of a system described by the differential equations Xi = Xi - X2 + X3
X2 = 2xi + 3X2 + -^3 ^3
= 3JCI + 2X2 +
2X3.
6.2. Determine the set of equilibria of a system described by the differential equations Xi
=
X2
xi sin — ,
when xi T^ 0,
0,
when xi = 0.
^2 = S
6.3. Determine the equilibrium points and their stability properties of a system described by the ordinary differential equation x = x(x-l)
(14.1)
by solving (14.1) and then applying the definitions of stability, uniform stability, asymptotic stability, etc. 6.4. Determine the set of equilibria and their stability properties of a system described by the ordinary differential equation X = (cost)x (14.2) by solving (14.2) and then applying the definitions of stability, uniform stability, asymptotic stability, etc. 6.5. Determine the set of equilibria and their stability properties of a system described by the ordinary differential equation X = (4tsmt-2t)x
(14.3)
by solving (14.3) and then applying the definitions of stability, uniform stability, asymptotic stability, etc. 6.6. Determine the state transition matrix ^(t, to) of the system Xi' X2_
-t 0 1 U\ _{2t - t) -2t\ [x2
Use Theorems 5.1 to 5.4 to determine the stability properties of the trivial solution of this system. 6.7. Show that the second-degree polynomial f{s) = s^ + 2as + b is a Hurwitz polynomial if and only if a> 0 and Z? > 0 by (i) solving the equation f{s) = 0, and (ii) using the Routh-Hurwitz criterion (Theorem 6.4). 6.8. Determine whether the third-degree polynomial f(s) = s^ + 3s^ + 3s + 2
511 CHAPTER 6: StabiHty
512 Linear Systems
is a Hurwitz polynomial by (i) solving the equation f(s) = 0, (ii) applying Theorem ^•^' ^^^^^ ^PP^yi^g Theorem 6.2, and (iv) applying Theorem 6.4. 6.9. Let A G C[/?+, /?"><"] and x G 7?" and consider X = A(t)x.
(LH)
Show that the equilibrium x^ = 0 of (LH) is uniformly stable if there exists a 2 G C^ [/?+, Z?^"""] such that 2 ( 0 = [Q(t)f for all r and if there exist constants C2 > ci > 0 such that cil < 2 ( 0 < C2/,
r G R,
(14.4)
and such that [A(0]^2(0 + Q(t)A(t) + m
< 0,
t^R,
(14.5)
where / is the fz X ^ identity matrix. Hint: The proof of this assertion is similar to the proof of Theorem 7.1. 6.10. Show that the equilibrium Xg = 0 of {LH) is exponentially stable if there exists a 2 ^ C^/?^,^''''''] such that 2 ( 0 = [2(01^ for alU and if there exist constants C2 > ci > 0 and C3 > 0 such that (14.4) holds and such that [A(0]^2(0 + Qit)A{t) + 2 ( 0 ^ -C3I,
t G R.
(14.6)
Hint: The proof of this assertion is similar to the proof of Theorem 7.2. 6.11. Assume that the equilibrium Xg = 0 of {LH) is exponentially stable and that there exists a constant a> 0 such that ||A(0|| ^ a for all t G R, Show that the matrix given by 2 ( 0 -^ j [a)(T,0]'^O(T,0^T
(14.7)
satisfies the hypotheses of the result given in Exercise 10. Hint: The proof of this assertion is similar to the proof of Theorem 7.5. 6.12. For {LH) let A^(0 and AM(0 denote the smallest and largest eigenvalues of A{t) + [A{t)]^ at r G /?, respectively. Let (/)(^, ^o, XQ) denote the unique solution of {LH) for the initial data x{to) = XQ = (^{t^, to, x). Show that for any XQ G R^ and any ^o G R, the unique solution of {LH) satisfies the estimate, 11^^11^(1/2)1,; A„(.).x ^ ||^(^^ ^^^ ^^^11 ^ ||^^||^o/2)(,; A,(.)..^ ^ ^ ^^^ ^j4 g^ //mL- Let v(r, ^0, XQ) = [4>{t, to, xo)V[(l){t, to, xo)] = \\(l){t, to, xo)p, evaluate v{t, to, xo), and then estabHsh (14.8). 6.13. Use Exercise 6.12 to show that the equilibrium Xg = 0 of {LH) is uniformly stable if there exists a constant c such that ft
XM{T)dT^c
(14.9)
for all t, a such that t > a, where AM(0 denotes the largest eigenvalue of A{t) + [A{t)Y, t G R. Hint: Use (14.8) and the definition of uniform stabihty. 6.14. Use Exercise 6.12 to show that the equilibrium Xg = 0 of (LH) is exponentially stable if there exist constants e > 0, a; > 0 such that XM{T)dT < -a{t
- (7) + 6
(14.10)
for all t, a such that t > a. Hint. Use (14.8) and the definition of exponential stabihty.
6.15. Let V be a quadratic function of the form V(XJ)
513
= x'^Q(t)x,
(14.11)
where xGR^'.Qe C^[R, /?"><«], Q(t) = [2(01^, and Q(t) ^ kl,k> 0, for all t e R. Evaluate the derivative of v with respect to t, along the solutions of {LH), to obtain ViLH){x, t) = x^[[A(0]^G(0 + Q(t)A(t) + Q(t)]x.
(14.12)
Assume that there is a quadratic form w(x) = x^Wx < 0, where W'^ = W G 7?«x«, such that V(LH)(x,t)^
W(X)
(14.13)
for all (x, t) G G X R, where G is a closed and bounded subset of /?". Let E = { x E G : w(jc) = 0}
(14.14)
and assume that for (LH), \\A(t)\\ is bounded on R. Prove that any solution of (LH) that remains in G for all ^ > fo — 0 approaches £" as ^ ^ oo. 6.16. Consider the system X + a(t)x + X = 0, which by letting xi = x and X2 - x can be written as Xi = X2 X2 = -a(t)X2
-
Xi.
Assume that a G C[R, R^] and that there are constants ci, C2 such that 0 < ci < a(t) ^ C2 for all t G R. Let v(x) = x\-\r x\. First, show that all solutions of this system are bounded. Next, use the results of Exercise 6.15 to show that ^2it, to, -^o) -> 0 as
6.17. Assume that for system {LH) there exists a quadratic function of the form v(jc, 0 = x^Q(t)x, where Q(t) = [Q(t)f G C^[R, 7?"''"] and Q(t) > cl for some c > 0, such that V(LH)(x, t) < x^Wx, where W = W'^ G R""^"" is negative definite. Show that if v is negative for some {x, t), then the equilibrium Xg = 0 of system {LH) is unstable. Hint. The proof of this assertion follows along similar lines as the proof of Theorem 7.3. 6.18. It is shown that if the equilibrium Xe = 0 of system {LH) is exponentially stable, then there exists a function v that satisfies the requirements of the result given in Exercise 6.10, i.e., the present result is a converse theorem to the result given in Exercise 6.10. In system {LH), let A be bounded for all ^ G /?, let L = L^ G C^[R, /?"><"] and assume that L is bounded for all t G R. Show that the integral Q{t)=
f
ma,t)VL{a)^{(T,t)da
exists for all t G R. Show that the derivative of the function v{xj)
= x^Q{t)x
(14.15)
with respect to t along the solutions of {LH) is given by y(LH){xj)
=
-X^L{t)x.
Next, show that if L{t) > c^I, C3 > 0, for all t G R, then there exist constants C2 ^ Cl > 0 such that for all t G R, c i / < Q{t)^ where / G /?"^" denotes the identity matrix.
C2I,
(14.16)
CHAPTER6:
Stability
514 Linear Systems
Note that the above result constitutes a generalization to Theorem 7.5 for timeinvariant systems (L). 6.19. Apply Proposition 7.1 to determine the definiteness properties of the matrix A given by 1 A = 2 1
2 5 -1
1 -1 10
6.20. Use Theorem 7.3 to prove that the trivial solution of the system 3 2
4 1
is unstable. 6.21. Determine the equilibrium points of a system described by the differential equation X = -X + x^ and determine the stability properties of the equilibrium points, if applicable, by using Theorem 8.1 or 8.2. 6.22. The system described by the differential equations Xi = X2 + Xi(xj + xl)
(14.17)
Xi + X2{x\ + x\) has an equilibrium at the origin x^ = {x\, X2) = (0, 0). Show that the trivial solution of the linearization of system (14.17) is stable. Prove that the equilibrium x = 0 of system (14.17) is unstable. (This example shows that the assumptions on the matrix A in Theorems 8.1 and 8.2 are absolutely essential.) 6.23. Prove that the system given by 'xx'
M.
{It
'yi
cost sint
J2J
-1)
0] \xi +
uit)
-2t\
sin^l U\ - cos^J [X2_
is BIBO stable. 6.24. Use Theorem 9.3 to analyze the stability properties of the system given by i: = Ax + Bu y = Cx A
1 1
0 -1
B =
1 -1
C = [0, 1].
6.25. Determine all equilibrium points for the discrete-time systems given by (a)
xi(k+
1) - X2(k) + \xi(k)\
X2(k+ 1) = -Xi(k) + \x2(k)\
(b)
Xi(k+
1) = Xi(k)X2(k)
515
- 1
X2(k + 1) = 2xi(^)jC2(^) + 1.
CHAPTER 6 :
Stability 6.26. Consider the discrete-time system given by x(k-{- 1) - sat[Ax(k)] where for 0 = (Oi,..., OnV e /?", sat6 = [satOi,..., r 5<2^^/
1,
(14.18) satdnV. and
^/ > 1,
=
Oi < 1.
-1,
(a) For A G /?"^" arbitrary, use Theorem 10.1 to analyze system (14.18). (b) Imposing various restrictions on the locations of the eigenvalues of A in the complex plane, use as many results of this chapter as you can to analyze the stability properties of the trivial solution of system (14.18). 6.27. Determine the stability properties of the trivial solution of the discrete-time system given by the equations cos 0 -sin^
'xx{k+ 1)"
Mk +1).
sin ^ 1 \xx{k) cosOj [x2{k)
with 6 fixed. 6.28. Analyze the stability of the equilibrium x = 0 of the system described by the scalarvalued difference equation x{k +1) = sin[x(^)]. 6.29 Analyze the stability of the equilibrium x = 0 of the system described by the difference equations Xx{k + 1) = Xx{k) + X2{k){Xi{kf
-h X2{kf^
X2(k + 1) = X2(k) - xi(k)[xi(kf
+
X2(kfl
6.30. Determine a basis of the solution space of the system xi(k+ 1) X2(k + 1)J
0 -6
1 xi(k) 5 [X2(k),
Use your answer in analyzing the stability of the trivial solution of this system. 6.31. Let A E /?"^". Prove that part (iii) of Theorem 10.8 is equivalent to the statement that all eigenvalues of A have modulus less than 1, i.e., lim ||Ai = 0 if and only if for any eigenvalue A of A, it is true that |A| < 1. 6.32. Use Theorem 10.7 to show that the equilibrium x = 0 of the system
x(k + 1) =
is unstable.
1 0
1 1
1 1
. . . .
1 1
0
0
0
. .
1
x(k)
516
6.33. (a) Use Theorem 10.9 to determine the stability of the equilibrium x = 0 of the system "1 x(k -\-l) = 0 0
Linear Systems
1 1
-21 3 x(k). 9 - 1
(b) Use Theorem 10.9 to determine the stabiHty of the equilibrium x = 0 of the system x(k+l)
ri = \o
0 1
-21 slxik).
[o 9
-ij
6.34. Apply the Schur-Cohn criterion (Theorem 10.10) in analyzing the stability of the trivial solution of the system given by the equations xi(y^+l)] r-0.5 X2(k+ 1) = 0.5
U3(/^+l)J
I. 0
0 0
0.5irxi(^) 0 \\x2(k)
-0.5
0JU3W,
6.35. Apply Theorems 7.2 and 10.11 to show that if the equilibrium x = 0 (x E. /?") of the system x(k +1) = e^x(k) is asymptotically stable, then the equilibrium x = 0 of the system X = Ax is also asymptotically stable. 6.36. Apply Theorem 10.11 to show that the trivial solution of the system given by xi(k+ 1) [X2(k + 1)J
0 [2
2 xi(k) Oj [X2(k)
is unstable. 6.37. Determine the stability of the equilibrium x = 0 of the scalar-valued system given by x(k+ I) = ^x(k) + f sin jc(^). 6.38. Analyze the stability properties of the discrete-time system given by x(k+ I) = x(k) + ^u(k) y(k) =
'^x(k)
where x, y, and u are scalar-valued variables. Is this system BIBO stable? Can Theorem 10.16 be applied in the analysis of this system?
CHAPTER?
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
In this chapter, representations of linear time-invariant systems based on polynomial matrices, called Polynomial Matrix Description (PMD) or Differential (Difference) Operator Representation (DOR) are introduced. Such representations arise naturally when differential (or difference) equations of order higher than one are used to describe the behavior of systems, and the differential (or difference) operator is introduced to represent the operation of differentiation (or of time-shift). Polynomial matrices in place of polynomials are involved since this approach is typically used to describe MIMO systems. Note that state-space system descriptions involve only first-order differential (or difference) equations, and as such, PMDs include the state-space descriptions as special cases. A rational function matrix can be written as a ratio or fraction of two polynomial matrices or of two rational matrices. If the transfer function matrix of a system is expressed as a fraction of two polynomial or rational matrices, this leads to a Matrix Fraction(al) Description (MFD) of the system. The MFDs that involve polynomial matrices, called polynomial MFDs, can be viewed as representations of internal realizations of the transfer function matrix, i.e., as system PMDs of special form. These polynomial fractional descriptions (PMFD) help establish the relationship between internal and external system representations in a clear and transparent manner. This can be used to advantage, for example, in the study of feedback control problems, leading to clearer understanding of the phenomena that occur when systems are interconnected in feedback configurations. The MFDs that involve ratios of rational matrices, in particular, ratios of proper and stable rational matrices, offer convenient characterizations of transfer functions in feedback control problems. 517
518 Linear Systems
MFDs that involve ratios of polynomial matrices and ratios of proper and stable rational matrices are essential in parameterizing all stabilizing feedback controllers. Appropriate selection of the parameters guarantees that a closed-loop system is not only stable but will also satisfy additional control criteria. This is precisely the approach taken in optimal control methods, such as //°°-optimal control. Parameterizations of all stabilizing feedback controllers are studied extensively in Part 2 of this chapter. We note that extensions of MFDs are also useful in linear time-varying systems and in nonlinear systems. These extensions are not addressed here. In addition to the importance of MFDs in characterizing all stabilizing controllers, and in //°°-optimal control, PMFDs and PMDs have been used in other control design methodologies as well (e.g., self-tuning control). The use of PMFDs in feedback control leads in a natural way to the polynomial Diophantine matrix equation that is central in control design when PMDs are used and that directly leads to the characterization of all stabilizing controllers. The Diophantine Equation is studied at length in Part 1 of this chapter. Finally, PMDs are generalizations of state-space descriptions, and the use of PMDs to characterize the behavior of systems offers additional insight and flexibility. These issues are also explored in Part 1 of this chapter.
7.1 INTRODUCTION In this chapter. Polynomial Matrix Descriptions (PMDs) and Matrix Fractional Descriptions (MFDs) are used to study properties such as controllability, observability, and stability, primarily of interconnected systems, and to conveniently characterize all stabilizing feedback controllers. These system descriptions are important in feedback control system analysis and design and are the key to developing control design theories such as H'^-opiimal control. The development of the material in this chapter is concerned only with continuous-time systems; however, completely analogous results are valid for discrete-time systems and can easily be obtained by obvious modifications. In the following, PMDs and MFDs are first introduced by an illustrative example. Next, the contents of the chapter are briefly described and some guidelines for the reader are provided. An important comment on notation In this chapter we will be dealing with matrices with entries that are polynomials in s or q, denoted by, e.g., D(s) or D(q). For simplicity of notation we frequently omit the argument 5* or ^ and we write D to denote the polynomial matrix on hand. When ambiguity may arise, or when it is important to stress the fact that the matrix in question is a polynomial matrix, the argument will be included.
A. A Brief Introduction to Polynomial and Fractional Descriptions The PMD and the MFD of a linear time-invariant system are introduced via a simple motivating example.
EXAMPLE 1.1. In the ordinary differential equation representation of a system given by yi(t) + yi(t) + yiit) = uiit) + ui(t) yi(t) + Ht) + 2y2(t) = U2(t)
(1.1)
yi(t\ yiit) and ui(t), U2(t) denote, respectively, outputs and inputs of interest. We assume that appropriate initial conditions for the w/(r), yi(t) and their derivatives at r = 0 are given. By changing variables, one can express (1.1) by an equivalent set of first-order ordinary differential equations, in the sense that this set of equations will generate all solutions of (1.1), using appropriate initial conditions and the same inputs. To this end, let xi = yi - U2,
X2 = yi,
(1.2)
yi + y i - U2.
X3
Then (1.1) can be written as y = Cx + Du,
X = Ax + Bu, lxi(t) where x{t) = X2(t)
Ui(t) U2(t)
u{t) =
X3(t)\
0 A = 1 0
0 0 2
-1 0 -2
1 B = 0 0
y(t) = -1 1 -2
yi(t) yiit).
(1.3)
and
C =
1 -1
D =
with initial conditions x(0) calculated by using (1.2). More directly, however, system (1.1) can be represented by P(q)z(t) = Q{q)u{t), where P(q)
Z{t) = 2+ 1 1 q q+2
zx{t) Z2(t)\
, U(t) =
Q(q)-
y{t) = R{q)z{t) + W{q)u{t), Ui(t)
uiit) q q
yi(t) y{t) = yiit). R(q)
(1.4)
and
I 0
W(q)
with q = didt, the differential operator. The variables zi{t), ziif) are called/7(2r^/(2/ state variables, z(t) denotes iht partial state of the system description (1.4), and u{t) and y{t) denote the input and output vectors, respectively. • Polynomial matrix descriptions (PMD) Representation (1.4), also denoted as {P{q), Q{q), R{q), W{q)}, is an example of a PMD of a system. Note that the state-space description (1.3) is a special case of (1.4). To see this, v^rite (1.3) as {ql - A)x{t)
= Bu{t\
y(t) = Cx(t) + Du(t),
(1.5)
Clearly, description {ql - A, B, C, D] given in (1.5) is a special case of the general P M D {P{q), Q(q\ R(q\ W(q)} with P(q) = ql-A,
Q(q) = B,
R{q) = C,
^{q)
= D.
(1.6)
The above example points to the fact that a P M D of a system can be derived in a natural way from differential (or difference) equations that involve variables that are directly connected to physical quantities. By this approach, it is frequently
519 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
520 Linear Systems
possible to study the behavior of physical variables directly without having to transform the system to a state-space description. The latter may involve (state) variables that are quite removed from the physical phenomena they represent, thus losing physical insight when studying a given problem. The price to pay for this additional insight is that one has to deal with differential (or difference) equations of order greater than 1. This typically adds computational burdens. We note that certain special forms of PMDs, namely, the polynomial MFD, are easier to deal with than general forms. However, a change of variables may again be necessary to obtain such forms. Consider a general PMD of a system given by P{q)z{t) = Q(q)u(t\
y(t) = R(q)z(t) + W(qMt)
(1.7)
withP(^) G Rlgy^'K Q(q) G Riq]^'''^ and R(q) G RlqV^'K W(q) G /?[^]^X'^, where Rlq^^^ denotes the set of / X / matrices with entries that are real polynomials in q. The transfer function matrix H(s) of (1.7) can be determined by taking the Laplace transform of both sides of the equation assuming zero initial conditions (z(0) = ^(0) = ... = 0, u(0) = u(0) = '" = 0). Then H(s) = R(s)p-\s)Q(s)
+ W(s).
(1.8)
For the special case of state-space representations, [see (1.6)]; H(s) in (1.8) assumes the well-known expression H(s) = C(sl — Ay^B + D. For the study of the relationship between external and internal descriptions, (1.8) is not particularly convenient. Indeed, it appears that it is as difficult to investigate the relationship between H{s) and PMDs as it was to study the relationship between H{s) and state-space descriptions. There are, however, special cases of (1.8) that are very convenient to use in this regard. In particular, as will be shown in Section 7.3, if the system is controllable, then there exists a representation equivalent to (1.7) that is of the form DM)Zcit)
= u{t),
yit) = NM)Zc{t\
(1.9)
where Ddq) G R[qY''''^ and Nc{q) G Riq^"^. Representation (1.9) is obtained by letting Q{q) = I^ and W(q) = 0 in (1.7). It is common to use D and N instead of P and R, in view of H(s) = Nc(s)Dc(sr\
(1.10)
where Nds) and Dds) represent the matrix numerator and matrix demonimator of the transfer function, respectively. Similarly, if the system is observable, there exists a representation equivalent to (1.7) that is of the form Do(q)Zo(t) = No(q)u(t),
y(t) = ZoU),
(1.11)
where Do{q) G R[q\P^P and A^^(^) G RiqY^"^. Representation (1.11) is obtained by letting R{q) - Ip and W{q) = 0 in (1.7) with P{q) - Do{q) and Q{q) = No{q)Here, H{s) = D-\s)No{s\
(1.12)
Note that (1.10) and (1.12) are generalizations to the MIMO case of the SISO system expression H{s) = n(s)/d(s). In the same manner as H(s) = n(s)/d(s) can be derived from the differential equation d(q)y(t) = n(q)u(t), (1.12) can be derived from (1.11), usually written as i)o(^)};(0 = No(q)u(t).
Returning now to (1.3) in the example, notice that the system is observable (state observable from the output y). Therefore, the system in this case can be represented by a description of the form {Do, No, h, 0}. In fact (1.4) is such a description, where 1 , and Do and No are equal to P and Q, respectively, i.e., Do(q) = q^+ 1 q+2
1 q . The transfer function matrix is,given by [O q C{sl- Ay^B + D
No(q) =
H(s)
' s 0 1 " "0 1 Ol - 1 ^ 0 0 - 1 ij 0 -2 s + 1 D-\s)No{s) = 1
52 + 1 5
\s + 2
^3 + 2^2 + 2 [ -s
1 " 5+ 2
- 1 r
-1
"1 - 1 "0 0" 1 + 0 1 0 0 -2_ "1 s 0 s
- 1 1 ri s s^ + ij [0 s =
S'
1 + 2^2 + 2
s{s + 1) -s
Matrix fractional descriptions (MFDs) of system transfer matrices A given pXm proper, rational transfer function matrix H{s) of a system can be represented as H{s)
= NR(S)D^\S)
=
DI\S)NL(S),
(1.13)
whevQ NR(S) E Rlsy'^.DRis) E R[sr'''^ mdNiis) E RlsV'^.DUs) E RISVP. The pairs {NR(S), DR(S)} and {DL(S), NL(S)} are called Polynomial Matrix Fractional Descriptions (PMFDs) of the system transfer matrix with {NR(S\ DR(S)} termed a right Fractional Description and {DL(S), NL(S)} a left Fractional Description. Notice that in view of (1.10), the right Polynomial Matrix Fractional Description (rPMFD) corresponds to the controllable PMD given in (1.9). That is, {DR, Im, NR, 0}, or DR(q)zR(t) = u(t),
y(t) = NR(q)zR(t)
(1.14)
is a controllable PMD of the system with transfer function H(s). The subscript c was used in (1.9) and (1.10) to emphasize the fact that Nc, Dc originated from an internal description that was controllable. In (1.13) and (1.14), the subscript R is used to emphasize that {NR, DR} is a right fraction representation of the external description H(s). Similarly, in view of (1.12), the left Polynomial Matrix Fractional Description (IPMFD) corresponds to the observable PMD given in (1.11). That is, {DL, NL, Ip, 0}, or DL(q)ZL(t) = NdqMt),
y(t) = ZLO)
(1.15)
is an observable PMD of the system with transfer function H(s). Comments analogous to the ones made above concerning controllable and right fractional descriptions (subscripts c and R) can also be made here concerning the subscripts o and L. An MFD of a transfer function may not consist necessarily of ratios of polynomial matrices. In particular, given a /? X m proper transfer function matrix H(s), one
521 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
522
can wnte
Linear Systems
H(s)
= NR(S)D^\S)
=
(1.16)
DI\S)NL(S\
where NR, DR, DU NL are proper and stable rational matrices. To illustrate, in the example considered above, H{s) can be written as H{s)
1
5 + 2
s^ + 2^2 + 2
-s
'\{S + 1)2
[
0
0 s+ 2
-1
s{s + 1) S{S^ - 5 + 1 )
?2 + 1
1 s+2
52 + 1
1
1
(5 + 1)2 5"
{S + 1)2
(^ + 1)2
1
0
\s + 1)2 0
{S + 1)2
=
0 5+ 2
DI\S)NL{S).
^+ 2 J L ^+2 Note that Di{s) and A^L(^) are proper and stable rational matrices. Such representations of proper transfer functions offer certain advantages when designing feedback control systems. They are discussed further in this chapter in Part 2, Subsection 7.4D.
B. Chapter Description This chapter consists of two principal parts. In Part 1, the emphasis is on properties of systems described by PMDs. First, background on polynomial matrices is provided in Section 7.2, and the Diophantine Equation is studied at length in Subsection 7.2E. Equivalence of representations and system properties are addressed in Section 7.3. Properties of systems consisting of subsystems interconnected in parallel, in series (cascade), and in feedback configurations are investigated in Subsection 7.3C. In Part 2, Section 7.4, feedback control systems are studied with emphasis placed on parameterizing all stabilizing feedback controllers. Further details follow. In Section 7.2, polynomial matrices and their properties are studied, and special forms for polynomial matrices, which are useful in subsequent developments, are introduced. In particular, polynomial matrices in column reduced, triangular, Hermite, and Smith form are defined, and algorithms to obtain such forms by pre- and postmultiplication by unimodular matrices are given in Subsections 7.2B and 7.2C. Coprimeness of polynomial matrices is related to controllability and observability of PMDs and is studied in Subsection 7.2D. The Diophantine Equation, which plays a central role in feedback control, is studied at length in Subsection 7.2E, and methods for deriving particular solutions are given. PMDs of systems are addressed throughout Section 7.3. Controllability, observability, and stability are revisited in Subsection 7.3B. Also, PMD realizations of transfer function matrices are studied and realization algorithms are developed. The relationships among different PMDs and state-space descriptions of a system are explored in Subsection 7.3A, using equivalence of representations. The properties of systems consisting of interconnected subsystems are best explored using PMDs, and this is accomplished in Subsection 7.3C.
Feedback control systems are studied in Section 7.4 using PMDs and MFDs with emphasis on stabilizing controllers. All stabilizing controllers are parameterized using PMDs, in Subsection 7.4A, and proper and stable MFDs in Subsection 7.4C. State feedback controllers and state observers, important in the development involving MFDs, are discussed in Subsection 7.4B. The relationships among all feedback controller parameterizations discussed herein are derived and fully explained. The complete theory of parameterizing all stabilizing feedback controllers is developed in this section. Two degrees of freedom controllers, their stability properties, and their parameterizations are explored in Subsection 7.4D. Several implementations of such controllers are introduced and their limitations are addressed. Finally, several control problems such as the model matching problem, the diagonal decoupling problem, and the static decoupling problem are formulated and briefly discussed.
C. Guidelines for the Reader As with every chapter of this book, this chapter can be approached at different levels. If the characterization of all proper stabilizing feedback controllers using proper and stable MFDs, which arises in optimal control problems, is of primary interest, then the reader should focus on Subsection 7.4C. For better understanding of such MFDs of systems and of polynomial MFDs and their use in the study of systems, the reader at first reading should study selected topics from all sections of this chapter. In the following, the material that should be covered at first reading is described. The reader should first study coprimeness of polynomial matrices in Subsection 7.2D with emphasis on the tests for coprimeness (Theorem 2.4). To determine a greatest common divisor of polynomial matrices, one needs the algorithms given in Subsection 7.2D. All solutions of the Diophantine matrix equation are derived in Theorem 2.15 of Subsection 7.2E with particular solutions obtained in Lemma 2.14. The study of equivalence representations in Subsection 7.3A leads to insight concerning the relationships between different PMDs and state-space descriptions of a system. Tests for controllability, observability, and stability are given in Theorems 3.4,3.5, and 3.6 of Subsection 7.3B. Feedback configurations of interconnected systems are studied in Subsection 7.3C. Here the closed-loop descriptions are also derived, which are then used in Section 7.4 to study the class of stabilizing feedback controllers. All stabilizing controllers are expressed in terms of PMDs in Theorem 4.1 of Subsection 7.4A. Different parameters are introduced in Theorems 4.2 and 4.3, and Corollaries 4.5, 4.6, and 4.9. To fully understand proper and stable MFDs of systems, one needs to study state feedback controllers and state observers in terms of PMDs. This is accomplished in Subsection 7.4B. All proper stabilizing controllers are parameterized (using proper and stable MFDs) in Theorem 4.13 given in Subsection 7.4C and also in Theorem 4.16. The exact relationship between such MFDs and internal descriptions is provided by Theorem 4.20. Two degrees of freedom controllers that offer advantages concerning attainable system responses are studied in Subsection 7.4D. Theorem 4.21 is the principal stability theorem with all stabilizing controllers being parameterized in Theorem 4.22.
523 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
524 Linear Systems
Theorem 4.23 characterizes the response maps attainable via two degrees of freedom controllers under internal stability. Several control configurations are then examined. Finally, several control problems are formulated.
PARTI ANALYSIS OF SYSTEMS 7.2 BACKGROUND MATERIAL ON POLYNOMIAL MATRICES Let R[s]P^^ denote the set of p X m matrices with entries that are polynomials in s with real coefficients. If P(s) G R[sy^^, then P(s) will be called a. p X m polynomial matrix. Frequently, it will be necessary to determine the rank of P(s), which is defined as the maximum number of linearly independent rows (or columns) of P(s) over the field of rational functions. The rank of a polynomial matrix is discussed in Subsection A. In Subsection B, unimodular matrices are introduced and transformations of polynomial matrices to column and row reduced form are discussed. Hermite and Smith canonical forms are addressed in Subsection C, and in Subsection D the important concept of coprimeness of polynomial matrices is studied. In Subsection E the linear Diophantine Equation is examined. A. Rank and Linear Independence The linear independence of a set of vectors in a vector space, defined in Chapter 2, is recalled here for convenience. Let (V, F) denote a vector space V over thefieldF, and let v/ G K / == 1,..., /c. The set of vectors {vi,..., Vyt} is F-linearly dependent, i. e., it is linearly dependent over thefieldF, if there exists a set {ai,..., a^^jof scalars in F with at T^ 0 for at least one /, such that a\v\ + fl2V2 + • • • + akVk = Oy.
(2.1)
The set of vectors {vi,..., v^^} is linearly independent over thefieldF if (2.1) implies that ai = 0 for each / = 1,..., k. Linear dependence of p X 1 polynomial vectors pi(s) G R[sy is defined similarly. This warrants some explanation. Let R[s] be the ring of polynomials with coefficients in R (see Subsection 7.2E) and let R(s) be the field of rational fractions over R[s] (called the field of rational functions), i.e., R(s) = {t(s)\t(s) = ^ , with n,dG R[sl d ^ 0}. d(s) Note that if p(s) G R[s], it can always be considered as being divided by 1, in which case p(s)/l is a scalar in the field of rational fractions over R[s]. Thus, a polynomial vector pi(s) G R[S]P may be viewed as a special case of a rational vector, i.e., Pi(s) G R(sy, whose elements are in thefieldof rational fractions. Now consider the polynomial vectors pi(s) G R[sy, i = I,..., k. The set of vectors {pi(s),..., Pk(s)} is said to be R(s)-lmQ2ir\y dependent, (i.e., linearly dependent over the field of
rational functions), if there exists a set {ai(s),..., a^is)} of rational functions, (i.e., ai{s) G R{s), i = I,.. .,k) with ai{s) ¥' 0 for at least one /, such that ax{s)px{s) + ••• + ak{s)p^{s)
- 0 G
R{sy.
(2.2)
This set of vectors is linearly independent over R(s) if (2.2) implies that aiis) = 0 for each / = 1,..., k. 's + 3 = r '^^\,P2(s) 0 (s + l)/(s + 3), a2(s) = -I satisfy (2.2) since
EXAMPLE 2.1. LQipiis)
ai(s)pi(s) + a2(s)p2(s)
s+ 1 . Note that a 1(5') 0
s+ 1 s + 3 's+ 1 + (-1) s+3 0 0
Therefore, the set {p\(s), P2(s)} is linearly dependent over the field of rational functions. It is of interest to notice that {p\(s), p2(s)} is linearly independent over the field of reals. In particular, if ai, a2 are restricted to be reals then (2.2) implies that ai = a2 = 0 (verify this). This stresses the importance of the particular field over which linear independence is considered (refer to Section 2.2 of Chapter 2). • It is not difficult to see that if the set {pi (s),..., Pk(s)} is linearly dependent, then (2.2) is also satisfied for some polynomials ai(s). To see this, simply multiply both sides of (2.2) by the least common multiple of the denominators of the rational functions ^1(5"),..., ak(s). This implies that linear dependence over R(s) can be tested merely by searching for polynomials ai(s) E R[s], not all zero, satisfying (2.2). To illustrate, consider: EXAMPLE 2.2. In Example 2.1, {pi(s), p2(s)} is linearly dependent over R(s) since s + 3' 's+ 1 + (-(s + 3)) (^+1) 0 0 DEFINITION 2.1. The normal rank of a polynomial matrix P(s) G /?[5]^^'" is the maximum number of linearly independent rows (or columns) over the field of rational functions R(s). • EXAMPLE 2.3. (i) rankPi(s) rank
s s+1
1 s+ 1
= rank
r^ + 1 0
^ + 31 0
= 1, and (ii) rankP2{s) =
2.
It can be shown that the (normal) rank of P{s) is also equal to the order of the largest order nonzero minor of P{s). EXAMPLE 2.4. (i) P\{s) in Example 2.3 does not have a second-order nonzero minor, since det P\{s) = 0, and therefore its rank is less than 2. The entries are the first-order minors and since nonzero entries exist, rankPi(s) = 1. (ii) P2(s) in Example 2.3 has a second-order nonzero minor since det P2(s) ^ c 2 . 1 ^ 0 . Therefore, rank P(s) = 2. Notice that if ^ = 1 or - 1 , then det P2(l) = det P2(-1) = 0. This is true because in this case, P2(l) [or P2(-1)] has only one linearly independent column (over the field of reals R) and rank P ( l ) = 1. Since this loss of rank (from 2 to 1) occurred for only special values of s, rank P(s) defined above is referred to as the normal rank of P(s), instead of just the rank of P(s), when there is ambiguity. Note that unless otherwise stated, the linear dependence of rows or columns of a matrix, considered
525 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
526 Linear Systems
for rank evaluation, is taken over the smallest field that contains the entries of the matrix. B. Unimodular and Column (Row) Reduced Matrices A polynomial matrix U(s) G R[sy^^ is called unimodular (or /?[5-]-unimodular) if there exists a U(s) G R[SY^P such that U(s)U(s) = Ip. This is the same as saying that U~^(s) = U(s) exists and is a polynomial matrix. Equivalently, U(s) is unimodular if det U(s) = a G R,a7^0. It can be shown that every unimodular matrix is a matrix representation of a finite number of successive elementary row and column operations. The elementary row and column operations on any polynomial matrix P{s) E R[sY^^ consist of the 1. interchange of any two rows (or columns) of P{s), 2. multiplication of any row (or column) of P{s) by a nonzero real a G 7?, a: 7^ 0 (a unit in R[sY), 3. addition to any row (or column) of P{s) of a multiple by a nonzero polynomial p{s) of another row (or column). These elementary row (and column) operations can be performed by multiplying P{s) on the left (right), [i.e., by pre-(post-) multiplying P{s)\, by elementary unimodular matrices. These elementary unimodular matrices are obtained by performing the elementary operations (1) to (3) on the identity matrix /. As mentioned above, it can be shown that every unimodular matrix may be represented as the product of a finite number of elementary unimodular matrices. EXAMPLE 2.5. The interchanging of rows 1 and 3 in a specific example is accomplished, e.g., as shown: "0 0 11 1 0 0 ^+2 UL{S)P{S) = 0 1 0 s + 1 1 s+ 1 1 _1 0 oj 0 ^+2 1 0
Also, addition to the second column of the third example is accomplished, e.g., as shown: s~\ri 0 1 0 P(S)UR(S) = 5+ 1 0 0 1 1 0 s + 2 ij LO s
column multiplied by 5" in a specific 0" 1 0 = s+ 1 1 2s + 2 1. 0
Let the degree of a polynomial (row or column) vector be the degree of the highest degree entry and let deg^. (P) [deg^. (P)] denote the degree of the /th row (jth column) of P(s). Also let Cr(P) [Cc(P)] be the highest row degree (column degree) coefficient matrix of P(s), defined as the real matrix with entries that are the coefficients of the highest degree s terms in each row (column) of P(s). We note that P(s) G RlsY^"^ can be written as P(s)
diag (s'^'i, s'^'-2 ^
Cc(P)diag(s'^
S--p)Cr(P) + Pr(s)
'^,S^'2,
wheredn = deg^. (P), i = 1,..., p, anddcj Pc(s) appropriate polynomial matrices.
..
.,S^'
) + Pc(sl
(2.3)
deg^. (P), j = 1 , . . . , m, with Pr(s),
's+l 3^2 + 21 . The row degrees are deg^^ (P) s 1 [5^ + 3 s^ + 5 , 2,deg^2(P) ^ 1' ^^^ ^^^r3 (^) ^ ^' while the column degrees are deg^i(P) = 2
EXAMPLE 2.6. Let P(s)
=
and J^^c2 (P) ^ 3- The highest row degree coefficient matrix of P(s) is Cr(P) 0 and the highest column degree coefficient matrix is CdP) = 0 1 0 0 0
P(s)
Cc
0
0
0 0
0 0
Cr + Pr{s)
Sdr3
+ Pc(s)
0 0 1
s 0
0 0 1
0 ^3 0
0 1 0
ro 3] 1 0 .0 1_
0 0 We have 1 3 0 1
+
s +1 s 3
s+l 0 s^ + 3
2 1 5
3^2 + 1 1 5
P(s) is row (column) proper, also called row {column) reduced, if Cr{P) has full rank.
[CdP)]
EXAMPLE 2.7. P{s) in Example 2.6 is row proper since rankCr = 2 but not column proper since rank Cp = I <2. • Consider now the first two rows of ^(5") in Example 2.6 and note that det
s +1 s
3s^ + 2 = (-3)/ 1 = det
0 1
s+ 1 3 ^(dri +dr2) _^ lower degree terms . 0
This illustrates the following result that can be derived from (2.3): any highest order minor of P{s) E 7^ [5']^^'" is a polynomial of degree equal to the sum of the degrees of the rows {p > m) or of the columns (;? < m), with leading coefficient equal to the corresponding minor of Cr{P) or of CdP), respectively. For the case when P(s) is square, this immediately implies that detP(s)
= detCdP)s^^'^
+ lower degree terms
= detCdP)s^^'^
+ lower degree terms.
(2.4)
Clearly then, P{s) G R[sY^^^ is row (column) proper if and only if ^i^^ {det P{s)) ^ X dr,{deg{det{P{s)) = Z dcj). (Show this.) Note that if P{s) G R[S]P'''^ is not of full rank it can be neither a column nor a row proper matrix. Equation (2.3) leads to useful polynomial matrix representations given by P{s) = diag [s'^'i^Cr + block diag
[l,s,, •.s'^^i-^]Cr (2.5a)
= block diag [1, ^ , . . . , s^'i]Pr and
P{s) = Cc diag [/^^ ] + Cc block diag [[1, s,,,,, =
Pcblockdiag{[\,s,...,s'^'jfl
s'^'j ^ ] (2.5b)
527 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
EXAMPLE 2.8. In view of (2.5a), we have the representation
528
r 1
Linear Systems
P(s)
s +1 s s^ + 3
s'
3s^ + 2"
1
= 0
s^ +5 j
0
0 01 ro 3" 1 5 0 0 0 0 1 0 + 0 0 1 0 0 0 ^^J LO ij 0 0 0 1 5 s
2 1
1
0
01 0 0
1
^'J 3
5 0 [ 1 0_ 0
1 0
21 0 3
0 1
1 0
r 1
2 : 0 0 1 s s^ 0 0 0 : 1 s 0 0 0 : 0 0
0 0 0 0 0 0 1 s
s'
0 0
^'J 3 0 1
N
5 0 0 1 _
Similarly, a representation in terms of column degrees can be obtained using (2.5b). • P{s) can also be expressed as a matrix polynomial of the form P{s) = Pks^ + Pk-is"-'
+ "' + Pis + Po,
(2.6)
where Pi G RP^^, A square polynomial matrix is called regular if rankP^ = m. Note that if P(s) is regular, then it is both row and column proper. EXAMPLE 2.9. P(5) of Example 2.8 can be written as P(5) = ro 01 ri 0" ri ro 3" s+ 1 3^2 + 2 = 0 0 ^3 + 0 1 s^ + 1 0 s + 0 1 0 1 0 0 1 0 3 s^ + 3
P3S^ + P2S^-\-Pis + Po = 21 1 5
Reduction to a row (column) proper polynomial matrix Given P(s) G Rlsy^'^ of full rank, there exists a unimodular matrix UL(S) such UL(S)P(S) is row proper. This is shown here by using a constructive proof. At each step of the algorithm below, the degree of a row (the highest degree row) is reduced by at least one, using elementary row operations. Since the matrix is of full rank, the algorithm will stop after a finite number of steps. that
Algorithm Let drf = degj.. (P), i = \,...,
p.
(i) Ohi2im diag[s^^i]Cr{P). (ii) Determine/? monomials pi{s) such that {p,,,..,Pp)diag[s'^qCr{P)
= Q.
Take pk = I. This is accomplished by dividing all monomials by the lowest degree (nonzero) monomial, assumed here to be the /cth one. (iii) Premultiply P(s) by
Ux(s) =
"1
0
•••
Pi
Pi
0
0
0
•••
0 Pp
0
kih row.
1
(iv) Stop if U\P is row proper. Otherwise, set P = U\P\ repeat the steps. To determine the appropriate unimodular matrix UL{S) directly, so that UL(S)P(S) is row proper, one may decide on the necessary row operations based on P(s) (as in the algorithm above), but apply these to [P(s\ Ip]. Then UL(S)[P(S), Ip] = [P(s), UL(S)], where P(s) is the row proper matrix and UL(S) can be read off directly from the resulting matrix. 's+ 1 EXAMPLE 2.10. Fov P(s) =
s dr, -
1,
and rankCr(P) = rank
s s^ + 2 , the row degrees are dr^ = I, dr^ = 2, s+2 = 1 < 2, and thus, P(s) is not row proper.
The algorithm is now applied. We have ' s 's 0 01 ri 1" (i) diag [s'^i][Cr(P)] = 0 s2 0 1 1 = s^ _s .0 C) ^ J Li 1. (nl(ni)Ui(s)
(iw)UiP =
ULP. where UL
s' s^ s_
" 1 0 0" = -s 1 0 ' . 0 0 1 . 's + 1 s ' -s 2 , aiidCriUiP) s s + 2.
U\, is row proper.
' 1 11 = - 1 0 , which is of full rank. Thus, 1 ij
•
Note that a unimodular matrix UL such that ULP is row proper is not unique. In [I 0 - 1 fact, in Example 2.10, another choice for UL could have been Ui = "1 - 2 " -2 1 , which is of full rank, and s^ + 2 with CriUiP) = 1 1 1 s+2 therefore, Ui P is row proper. It is possible to reduce a polynomial matrix to a row proper polynomial matrix using elementary column operations (in place of elementary row operations). In particular, given P(s) G R[sy^^ of full rank, there exists a unimodular matrix UR(S) such that P(S)UR(S) is row proper. Such a UR(S) is determined, for example, by the algorithm described in Subsection 7.2C, which reduces a matrix to a lower left Hermite form. Other algorithms to accomplish this can also be derived.
since U\P =
529 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
530
Linear Systems
EXAMPLE 2.11. Consider P(s) of Example 2.10 and apply the algorithm of Subsection 7.2C to reduce P(s) to lower left Hermite form. Take P(sY and determine U^(s) such that U^P^ is reduced to upper right Hermite form. Then P{s)
—s 1 = P(s) 1. -I
1 01 n . - 1 Ij
= P(S)UR(S)
-si s+l\ ' 1 0 = - 2 s^ + 2s + 2 .-2 3s+ 2
which is row proper. Similar results for reducing a polynomial matrix to a column proper polynomial matrix can easily be derived. Given P(s) E R[sy^^ of full rank, there exists a unimodular matrix UR(S) such that P(S)UR(S) is column proper. [Take P(sy and apply steps (i) to (iii) of the above algorithm.] Also, there exists a unimodular matrix UL(S) such that UL(S)P(S) is column proper. [Use the algorithm to reduce P(s) to upper right Hermite form described in Subsection 7.2C.] Finally, we note that if P(s) E R[sy^^ has full rank, there exist unimodular matrices UL and UR such that ULPUR is both row and column proper. One such example is when UL and UR are chosen so that ULPUR is in Smith form (see Subsection 7.2C). Proper rational matrices Recall that a rational matrix H(s) E R(sy^^ limH(s)
= D,
is called proper if DGRP'''^
5—»oo
and if D = 0, then H(s) is called strictly proper Frequently, H(s) is expressed as H(s) =
N(s)D-\sl
where A^(^) and Z)(^) are polynomial matrices (A^(^) E RlsV"^ Sind D(s) E Rls]"^"""^), The pair {N(s), D(s)} can be viewed as an rPMFD of a system described by a transfer function matrix H{s). It is of interest to relate the propemess of the rational matrix H(s) to the (column) degrees of N{s) and D(s). Note that when N(s) and D(s) are polynomials, it is easily seen that H(s) is strictly proper (is proper) if and only if degN(s) < degD(s) [if and only if degN(s) < degD(s)]. In the matrix case, similar necessary and sufficient conditions, given in Lemma 2.2, exist only when D(s) is column reduced. Necessary conditions for propemess are given in Lemma 2.1. Completely analogous results to Lemmas 2.1 and 2.2 hold for left factorizations of H(s) = D-\s)N(s) as well LEMMA 2.1. Let H(s) be a proper (or strictly proper) rational matrix and let H(s) = N(s)D(s)-\ Then deg^.N(s) < deg^.D(s) [or deg^.N(s) < deg^.D(s)] for j = 1,.... m. Proof. N(s) = //(5)D(^) and for the yth column of A^(^), m
^O'W = ^Hik(s)dkj(s\
i = I,.
•A
^=1
where nij(s) denotes the ith element in the jih column of N(s). Since every element Hik(s) of H(s) is strictly proper (or proper), all entries nij(s) must have degrees less than (or less than or equal to) the degree of the highest degree polynomial in the jih column of D(s). •
The converse to the above result is not always true. For example, let N{s) = [l,s]
and
D(s) =
s^ + \ s+l
CHAPTER?:
IJ
where deg^^ Nis) = 0 < deg^^ D{s) = 2 and deg^^ N(s) = 1 = deg^^ D(s). Here Nis)D-\s)
= {\,s]
2
1 1 -5 ? - 1 ^2 + 1 1 -s
-s^ — s 1 -^
1 ^ 1 -^
which is not proper. Notice that D{s) is not column reduced. When D{s) is column reduced (column proper) the conditions of the above lemma are sufficient as well, as the following result shows. LEMMA 2.2. Let H{s) = N(s)D~\s) with D(s) column reduced. Then H(s) is proper if and only if deg N(s) < deg,^ D(s),
j=
I
., m.
H(s) is strictly proper if and only if deg^. N(s) < deg^. D(s), j = 1,..., m. Proof, Necessity was shown in the previous lemma. To show sufficiency, notice that by applying Cramer's rule for the inverse to solve H(s)D(s) = N(s), we have hij(s) = [det D^J(sydet D(s)], where D^J(s) is the matrix obtained by replacing the jth row of D(^) by the /th row of N(s). In view of (2.3), D(s) can be written as D(s) = Cc(D) diag [s ""J ] + Dc(s), j = 1,..., m, where dcj = deg^. D{s\ CdD) is the highest column degree coefficient matrix of D(sX and deg^. D^s) < dc.. Similarly, D'^{s) = Cc(D'J) diag [ A ] + &-^(s), where Cc(D^^) is the same as CdD), except for the 7th row, which may or may not be zero since each entry of the jth row of N(s) is of lower than or equal degree of the corresponding entry of the jih row of D(s). Since D(s) is column proper, det CdD) ¥" 0, and in view of (2.4), degdetD(s) = X%idj while degdetD'ds) < YJ]=idj. Therefore, hij{s) is proper. It is strictly proper when deg^. N(s) < deg^. D(s) for j = 1,..., m. • EXAMPLE 2.12. H(s) =
D(s) =
\s + l
N(s)D~\s),whQrQ
-1
and
A^(^)=
ais + aQ CiS +
531
CQ
d\S -\r do
is proper for any values of the parameters since D(s) is column reduced and deg^. N(s) < deg^. D(s), j = 1, 2. H(s) is strictly proper only when ai = bi = ci = di = 0. • Let H{s) = D-\s)N(s), where D{s) G R[syp and N(s) G Rlsy"^. Completely analogous results to Lemmas 2.1 and 2.2 hold also for the left factorization matrices. In particular, if H(s) is proper (strictly proper), then deg^. N(s) < deg^. D(s) [deg^. N(s) < deg^.. D(s)] for / = 1 , . . . , /? (see Lemma 2.1). When D(s) is row reduced, then the conditions are necessary and sufficient (see Lemma 2.2).
C. Hermite and Smith Forms By elementary row and column operations, a polynomial matrix P(s) can be reduced to the Hermite form or the Smith form. These special forms are studied in this subsection, together with algorithms to reduce P(s) to such forms.
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
532
Linear Systems
Hermite form Given P(s) G R[sY^^ with p > m, there exists a unimodular matrix UL(S) such that UL(S)P(S) is an upper (right) triangular matrix of the form
UL(S)P(S)
Pm(s)
•• ••
X X
0
X X
0
0
0
0
••
0
0
0
••
0
"X
X
0
(2.7)
where Pm(s) G R[s]^^^. When p > m^ r = rankP(s), the last p - r rows are identically zero. In column J, 1 < j < r, the diagonal element is monic and of higher degree than any (nonzero) element above it. If the diagonal element is one, then all elements above it are zero. No particular form is assumed by the remaining m - r columns in the top r rows. This is the (upper triangular) column Hermite form. Note that if P(s) is of full rank, then UL(S)P(S) is column proper. By postmultiplication by a unimodular matrix UR(S), it is possible to obtain the row Hermite form of P(s) when p < m. To accomplish this, simply determine UL(S) in such a manner that UL(S)P^(S) is in column Hermite form, and take (UL(s)P^(s)y = P(s)Ul(s) = P(S)UR(S), which is in row Hermite form. The following algorithm reduces P(s) G R[sy^^, p > m, to (upper triangular) column Hermite form by elementary row operations [premultiplication by UL(S)]. This algorithm can also be used to constructively prove our desired result, that there exists a unimodular matrix UL(S) such that UL(S)P(S) (p > m) is in column Hermite form. The algorithm is based on polynomial division. Given any two polynomials a(s), b(s), b(s) # 0, there exist unique polynomials q(s), r(s) such that a(s) = q(s)b(s) + r(s) [q(s) is the quotient and r(s) is the remainder], where either r(s) = 0 or deg r(s) < deg b{s). By row interchange, transfer to the (1, 1) position the lowest degree element in the first column and call this element pn. Every other element pn in this column can be expressed by polynomial division, as a multiple of pn plus a remainder term of lower degree than pu, i.e.. qnPn + ni,
where deg rn < degp\\.
(2.8)
By elementary row operations, the appropriate multiple of j^n can be subtracted from each entry of column 1 leaving only remainders rn of lower degree than pi\. Repeat the above steps until all entries in the first column below (1, 1) are zero and note that the (1, 1) entry can always be taken to be a monic polynomial. Consider next the second column and position (2, 2) while temporarily ignoring the first row. Repeat the above procedure to make all the entries below the (2, 2) entry equal to zero. If the (1, 2) entry does not have lower degree than the (2, 2) entry, use polynomial division and row operations to replace the (1, 2) entry by a polynomial of lower degree than the (2, 2) entry. If the (2, 2) entry is a nonzero constant, use row operations to make the (1, 2) entry equal to zero. Continuing this procedure with the third, fourth, and higher columns results in the desired Hermite form.
533
EXAMPLE 2.13. s(s + 2) 0 P(s) = (s + l)(s + 2) 0
U3
t/6
where
5+2 0 0 0
0 (S+ 1)2 5+1 s(s + 1)
5+1 (5 + 1)2 -s(s + 1) s(s + 1)
5+2 5+ r 0 5+ 1 0 0 0 0
Uj
s(s + 2) 0 5+2 0
Ui
5+2 0 0 0
^4
"5 + 2 0
0 (s+ 1)2 5+1 5(5 + 1)
5+ 1 (s + 1)2 5+ 1 5+ 1
^5
U2
5+2 0 5(5 + 2) 0
5+2 0 0 0
CHAPTER?:
5+1 (5 + 1)2 0 S(S + 1)
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
5+1 5+1 (5 + 1)2 5+ 1
0 5 -f
1 =
0 0
0 0
0 0 1 0
0]ri
UL(S)P(S),
L(5) = U,U6'"Ui 1 0 0 0
-1 1 0 0
0 0
0 0
ij [0
0 1 -(5+1) -1
- ( 5 + 2) 5+1
-1 1
5+ 1
- ( 5 + 1)2
-5
S(S + 1)
-(5+I)
0 0" 0 0 1 0 0 1
' 1 0 0 0 0 1 0 0 - 1 0 1 0 0 0 0 1
-5
0
5
Notice that P{s) has full rank and UL{S)P(S) here is column proper.
•
Note that to determine UL(S) directly, one may decide on the necessary operations based on P(s), but apply these elementary row operations to [P(s),Ip]. Then UL(S)[P(S), Ip] = [Hp(s), UL(S)], where Hp(s) is the column Hermite form of P(s), (p > m). Also, note that if the algorithm is applied to ^(5") G R[S]P^^, where p ^ m, then
UL(S)P(S)
where P i (5) G
= [Pi(s\P2(s)]
Ris^P.
=
X
X
0
X
0
0
X
X
X
X
X
X
(2.9)
Smith form Given P(s) G R[sy^^ UL(S) and UR(S) such that
with rankP(s) UL(S)P(S)UR(S)
where
Sp(s) =
K{s) 0
0 0
A(^) =
= r, there exist unimodular matrices = Sp(s), diag{ei{s),.,.,er{s)).
(2.10)
534
Linear Systems
Each ei(s), i = 1 , . . . , r, is a unique monic polynomial satisfying ei(s) \ ei+i(s), i = 1 , . . . , r - 1, where P2(s)\pi(s) means that there exists a polynomial P3(s) such that Pi(s) = P2(s)p3(s), that is, €i(s) divides ei+i(s). Sp(s) is the Smith form ofP{s) and the ei{s), i = 1 , . . . , r, are the invariant polynomials of P{s). It can be shown that etis) =
Di(s) Di-i(sy
1, . . . , r ,
(2.11)
where Di(s) is the monic greatest common divisor of all the nonzero zth order minors of P(s). Note that Do(s) = 1 and Di(s) are the determinantal divisors of P(s). The Smith form Sp(s) of a matrix P{s) is unique, however, UL(S), UR(S) such that UL(S)P(S)UR(S) = 5'p(5') are not unique.
EXAMPLE 2.14. P(s)
Do = hDi
= hD2
=
s(s + 2) 0 0 (s+ 1)2 I with rankP(s) = r = 2. Here I (s + 1)(^ + 2) ^ + 1 0 s(s + 1)J
= (s + l)(s + 2),
and
ei
= DI/DQ
= 1, €2 = D2/D1
=
(s + l)(s + 2). Therefore, the Smith form of P(s) is
AW
1 0
0 (s+ l)(s + 2)
Sp(s)
0
The invariant factors 6/ of a matrix are not affected by row and column elementary operations. This follows from the fact that the determinantal divisors Dt are not affected by elementary operations (refer to the Binet-Cauchy formula in Exercise 7.3). In view of this, the following result can now be easily established: given Pi(s)> Piis) ^ R[s]P^^, there exist unimodular matrices Ui(s), U2(s) such that U,(S)PI(S)U2(S)
=
P2{S)
if and only if Pi(s), P2(s) have the same Smith form. The following algorithm reduces P(s) E R[sy^"^ to its unique Smith form Sp(s) and determines UL(S), UR(S) such that UL(S)P(S)UR(S) = Sp(s). The algorithm can be used to constructively prove that the Smith form of a matrix exists. Using row and column elementary operations, transfer the element of least degree in the matrix P(s) to the (1, 1) position. By elementary row operations, make all entries in the first column below (1, 1) equal to zero (refer to the algorithm for the column Hermite form). Next, by column operations, make all entries in the first row zero except (1, 1). If nonzero entries have reappeared in the first column, repeat the above steps until all entries in the first column and row are zero except for the (1, 1) entry. [Show that at each iteration the degree of the (1, 1) element is reduced, and thus the algorithm is finite.] If the (1, 1) element does not divide every other entry in the matrix, use polynomial division and row and column interchanges to bring a lower degree element to the (1, 1) position. Repeat the above steps until all other elements in the first column and row are zero and the (1, 1) entry divides every other entry in the matrix, that is.
eiis) 0
535
0
CHAPTER?:
Ei(s) 0 where 61(5') divides all entries ofEi(s). Repeat the above steps on Ei(s) and on other such terms, if necessary to obtain the Smith form of P(s). Two polynomial matrices Pi(s), P2(s) E R[sy^^ are said to be equivalent if there exist unimodular matrices U\{s), U2(s) such that Ui(s)Pi(s)U2(s)
= P2(S\
(2.12)
Recall that, as was mentioned earlier, (2.12) is satisfied if and only if Pi (s) and P2(s) have the same Smith form. The Smith form of P{s) is a canonical form for the relation (2.12) on R[sy^^. To see this, we first recall the definition of relation and equivalence relation: a relation p in a set X is any subset of X X X and p is an equivalence relation if and only if it satisfies the following axioms: 1. xpx-reflexivity (every x G X is equivalent to itself). 2. {xpy) =^ (3;px)-symmetry {x is equivalent to y implies that y is equivalent to x). 3. {xpy) and {ypz) =^ (xpz)-transitivity {x is equivalent to y and y is equivalent to z imply that x is equivalent to z). The relation p described by U1P1U2 = P2 in (2.12), (i.e., P1PP2), satisfies the above axioms (verify this) and therefore it is an equivalence relation. The pequivalence class or the "orbit" of a fixed P(s) G RlsY^'^ is denoted by [P(s)]p. The Smith form Sp(s) of P(s) is a canonical form for p on R[sY^^. D. Coprimeness and Common Divisors Coprimeness of polynomial matrices is one of the most important concepts in the polynomial matrix representation of systems since it is directly related to controllability and observability (see Subsection 7.3B). A polynomial g(s) is a common divisor (cd) of polynomials pi(sX P2(s) if and only if there exist polynomials pi(s), p2(s) such that Pi(s) = pi(s)g(s),
p2(s) = p2(s)g(s),
(2.13)
The highest degree cd of pi(^), P2(s), g'^is), is a greatest common divisor (gcd) of pi(s), P2{s). It is unique within multiplication by a nonzero real number. Alternatively, g*(5') is a gcd of pi{s), P2(s) if and only if any cd ^(^5") of pi(s), P2(s) is a divisor of g'^is) as well, that is. g^'is) = m(s)g(s)
(2.14)
with m(s) a polynomial. The polynomials pi(s), P2{s) are coprime (cp) if and only if a gcd g^'is) is a nonzero real. The above can be extended to matrices. In this case, both right divisors and left divisors must be defined, since in general, two polynomial matrices do not commute. Note that one may talk about right or left divisors of polynomial matrices only when the matrices have the same number of columns or rows, respectively.
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
536 Linear Systems
An mx m matrix GR{S) is a common right divisor (crd) of the pi x m polynomial matrix Pi{s) and the p2Xm matrix P2{s) if there exist polynomial matrices PiR{s),P2R{s)sothsit PI{S)=PIR{S)GR{S),
(2.15)
P2{S)=P2R{S)GR{S).
Similarly, a. p x p polynomial matrix GL{S) is a common left divisor (eld) of the pxmi polynomial matrix Pi {s) and the pxm2 matrix P2{s), if there exist polynomial matrices PIL{S)^P2L{S)
SO that
Px{s) = GL{S)PIL{S),
P2{S) =
(2.16)
GL{s)P2ds).
Also G^{s) is a greatest common right divisor (gcrd) of Pi{s) and P2{s) if and only if any crd GR{S) is an rd of G^(^). Similarly, G£(^) is a greatest common left divisor (geld) of A [s) and P2{s) if and only if any eld GL{S) is an Id of G£(^). That is. GI{S)=M{S)GR{S),
Gl{s)
=
(2.17)
GL{S)N{S)
with M{s) and N{s) polynomial matrices and GR{S) and Gi{s) any crd and eld of Pi {s), P2 ('^), respectively. Alternatively, it can be shown that any crd G\{s) of Pi{s) and P2{s) [or a eld G£(^) of A('^) and P2{s)] with determinant of the highest degree possible is a gcrd (geld) of the matrices. It is unique within a premultiplication (postmultiplication) by a unimodular matrix. Here it is assumed that GR{S) is nonsingular. Note that if rank : m [a (pi + P 2 ) X mmatrix], which is a typical case in polynomial matrix Piis) system descriptions, then rank GR{S) = m, that is, GR{S) is nonsingular. Notice also \Piis)\ that if rank GR{S) < m, then in view of Sylvester's Rank Inequality, rank <m Piis) as well. The polynomial matrices Pi (s) and P2 (s) are right coprime (re) if and only if a gcrd G^{s) is a unimodular matrix. Similarly, Pi{s) and P2{s) are left coprime (Ic) if and only if a geld G2{s) is a unimodular matrix. E X A M P L E 2.15.
Let Pi
5(5 + 2)
0
0
(5+1)2
Two distinct crds are GR^ =
0 5+1
and GR^
• ( 5 + l ) ( 5 + 2)
and P2 =
0
5+ 2
0
5(5+2)
0
0
0
5+1
(5+1)2
( 5 + l ) ( 5 + 2) 0
1
5+2
5+1
5
0
5(5+1)
GR.
A gcrd is G^
5+ 2
0
5+ 2
0
0
5+1
0
1
GR^
5+1 • 5(5+1)
G/?2
1
0
0
5+1
GR^.
NOW,
-»*—1
0
s 0
5+1
5+1
1
0
5
where Pf^ and P2j^ are re. Note that a geld of Pi and P2 is
G;
=
0 s+ 1
Both G^ and Gl were determined using an algorithm to derive
the Hermite form of
CHAPTER?:
as will be described later in this section.
It can be shown that two square p X p nonsingular polynomial matrices with determinants that are prime polynomials are both re and Ic. The converse of this is not true, that is, two re polynomial matrices do not necessarily have prime determinant polynomials. A case in point is Example 2.15, where P j ^ and PJ^ are re; however, detP\j^ = detP2R = s(s + 1). Left and right coprimeness of two polynomial matrices (provided that the matrices are compatible) are quite distinct properties. For example, two matrices can be Ic but not re, and vice versa (refer to the following example).
0 \s(s + 2) (s + 1)(^ + 2) 1 . J and P2 = I ^ I are Ic but not s+ 1. 0 s L 0 's + 2 0 with detGl = (s + 2). 0 1.
EXAMPLE 2.16. Pi = | ^ re since a gcrd is G^
537
Finally, we note that all the above definitions apply also to more than two polynomial matrices. To see this, replace in all definitions P i , P2 by P i , P2, • . . , P^. This is not surprising in view of the fact that the p\X m matrix P\{s) and the p2 X m matrix P2(s) consist of pi and p2 rows, respectively, each of which can be viewed as a 1 X m polynomial matrix; that is, instead of, e.g., the coprimeness of P i and P2, one could speak of the coprimeness of the (pi + P2) rows of P i and P2. How to determine a greatest common right divisor (gcrd) LEMMA 2.3. Let Pi(s) e /?M^i><^ and P2(s) e R[S]P^'''^ with pi + P2 ^ m. Let the unimodular matrix U(s) be such that U(s)
(2.18)
0
Piis)
Then G^(^) is a gcrd of Pi(^), P2(s). Proof. Let U =
X -Pi
Y Pi
(2.19)
withX G RisT""^^, y e R[sT''P\ P2 G RWP\2indPi G PM^><^2, where^ = (pi + P2) - m. Note that X, Y and P2, Pi are Ic pairs. If they were not, then det U 7^ a, 3. nonzero real number. Similarly, X, P2 and Y, Pi are re pairs. Let [/-'
\Pi
-Y]
(2.20)
vh ^\'
where Pi e RW^'^"', P^ G RisY^^"" are re and X G R[s]P2'"i, Y G RisV^'^'' are re. Equation (2.18) implies that
f/-'
0
Pi 1P2
Gl,
(2.21)
i.e., G^ is a crd of Pi, Pi- Equation (2.18) implies also that XPi + YP2 = G^.
(2.22)
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
538 Linear Systems
This relationship shows that any crd GR of Pi, P2 will also be an rd of G\. This can be seen directly by expressing (2.22) as MGR = G*j^, where Af is a polynomial matrix. Thus, G^ is a crd of Pi, P2 with the property that any crd GR of Pi, P2 is an rd of G^. This impHes that G^ is a gcrd of Pi, P2. • EXAMPLE 2.17. Let Pi =
's(s + 2) . 0
0 ,P2 (s + 1)^
(s + l)(s + 2) 0
^+ 1 ' s(s + 1)
Then
U
Pi
X -Pi
s+2 0
-(s + 2)
-1
s+ 1
0
s+ 1
1
-s
0
Y
Pi Pi} IP2I
0 s+ 1
-(s+lf
—s
s(s + 1)
-(s + 1)
0
s
Pi P2}
0 -1
G-
0 0 In view of Lemma 2.3, G^ =
5+ 2 0
0 s-hl
is a gcrd (see also Example 2.15).
Note that to derive (2.18) and thus determine a gcrd G^ of P i and P2, one could use the algorithm that was developed above, in Subsection 7.2C (or a variation of this algorithm) to obtain the Hermite form. Finally, note also that if the Smith form diaglEi] 0 of Pi is known, i.e., UL PI UR = SP = , then (diag [e/], 0)t/^ ^ is 0 0 Pi PI a gcrd of P i and P2 in view of Lemma 2.3. When rank Pi = m, which is the case Pi of interest in systems, then a gcrd of P i and P2 is diag [6/]L^^ ^. Criteria for coprimeness There are several ways of testing the coprimeness of two polynomial matrices, as shown in the following theorem. THEOREM 2.4. Let Pi E R[sV^'''^ and P2 G P[5]^2xm with pi + p2 ^ m. The following statements are equivalent: (i) Pi and P2 are re. (ii) A gcrd of Pi and P2 is unimodular. (iii) There exist polynomial matrices X G R[S]'^^P^ and Y G R[S]'^^P^ such that XPl + FP2 = Im. (iv) The Smith form of (v) rank
PliSi) PliSi).
m for any complex number Si.
constitutes m columns of a unimodular matrix.
(2.23)
Proof, Statements (i) and (ii) are equivalent by definition. Assume now that (iii) is true. Then (2.23) impUes that any crd of Pi and P2 must be an rd of Im, which is of course a crd of Pi and P2. Therefore, in view of the definition of a gcrd, Im is a gcrd (see also proof of Lemma 2.3) and so (ii) is true. To show that (ii) also implies (iii), determine a gcrd GJ as in the Lenrnia 2.3 and note that GJ is unimodular. Premultiplying (2.22) by we obtain (2.23). To show (iv), recall that a gcrd of Pi and P2 can be determined from the Smith form Px' of as GJ = (diag [£/], 0)t/^ ^ (refer to the discussion following Example 2.17). It is P2
now clear that a gcrd will be unimodular if and only if the Smith form of
IS
r/1 (i.e.. LOJ
\Pl(Si)] ^ [Piisi)! rank GJ(5'/). The only Si that can reduce the rank are the zeros of the determinant of GJ. Such Si do not exist if and only if detG]^ = a, a nonzero real number, i.e., if and only if G^ is unimodular. Therefore (v) implies and is implied by (ii). Part (vi) was shown in the proof of Lemma 2.3. • (iv) and (ii) are equivalent). To show (v), consider (2.18) and note that rank
0 s+ 1 P2 = ^+ 1 0 (see also Example 2.15) are re in view of the following relations. To use condition (ii) of Theorem 2.4, let
EXAMPLE 2.18. (i) The polynomial matrices Pi =
-(s + 2) U
s+ 1
Pi Pil
-1
s+ 1
1
\PI
[Pl.
-(s + If
—s
-(s + 1)
0
' 1 0
0" 1
0 0
0 0
=
=
s(s + 1)
. 0 .
Then G^ = h, which is unimodular. Applying condition (iii), XPi + FP2 =
" (^ + 2) -11 p + 1 0" s+1 l| ^ I -^ 0
/2.
To use (iv), note that the invariant polynomials of
Pi
areei = e2 = 1; the Smith
form is then h . To use condition (v), note that presently, the only complex values st 0 ~Pl(Si)] that may reduce the rank of are those for which detP\{Si) or detP2(si) = 0, P2(Si)\
i.e., ^1 = 0 and S2 = - 1 . For these values we have rank
and rank
'Pl(S2)
Piisi).
'-1 0 = rank 0 0
'PliSl) Piisi).
0 0 = rank 1 0
0" 0 = 2, i.e., both are of full rank. 1 -1
0 1 = 2 1 0
Note that if the criteria for coprimeness given in Theorem 2.4 are used for a pair P i , P2 that is not re, then (ii) will provide a gcrd G^ of P i , ^2- The other tests provide only partial information about G^. In particular, applying (iv), one obtains the Smith form of G^ (show this), while (v) will give information about the zeros of
539 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
540 Linear Systems
THEOREM 2.5. Let Pi G RlsV""^ and P2 ^ RisV"^^ with mi + m2 > p. The following statements are equivalent: (i) Pi and P2 are Ic. (ii) A geld of Pi and P2 is unimodular. (iii) There exist polynomial matrices X G R[S]'^I^P and Y G R[S]'^^^P such that PiX + P2Y = Ip.
(2.24)
(iv) The Smith form of [Pi, P2] is [/, 0]. (v) rank [Pife), P2(si)] = p for any complex number si. (vi) [Pi, P2] are p rows of a unimodular matrix. Proof. The proof is completely analogous to the proof of Theorem 2.4 and is omitted.
E. The Diophantine Equation The linear Diophantine Equation of interest to us is of the form X(s)D(s) + Y(s)N(s) = Q(sl
(DIO)
where D(s), N(s), and Q(s) are given polynomial matrices and X(s), Y(s) are to be determined. This equation will be studied in detail in this subsection. First, particular solutions of the Diophantine Equation are derived. Next, all solutions are conveniently parameterized. These parameterizations are used later in this chapter (in Section 7.4) to characterize all stabilizing linear feedback controllers of a linear time-invariant system. This subsection is concluded with some historical remarks. In greater detail, solutions of the polynomial Diophantine Equation of low degree are derived in Lemma 2.6, using the Division Theorem for polynomials, and in Lemma 2.8, using the Sylvester Matrix of two polynomials. Corresponding results for the polynomial matrix Diophantine Equation are derived in Lemma 2.12, using the Division Theorem and in Lemma 2.14, using the Eliminant Matrix. All solutions of the Diophantine Equation are parameterized in Theorem 2.15. The polynomial case Given coprime polynomials d{s) and n(s), it was shown in Theorem 2.4 that there exist polynomials x{s) and y{s) such that x(s)d(s) + y(s)n(s) = 1.
(2.25)
It can be shown that x(s) and y(s) are unique when certain restrictions are imposed on their degrees. In particular, the following result is true. LEMMA 2.6. Let d(s) and n(s) be coprime polynomials. Then there exist unique polynomials x(s) and 5^(^) that satisfy x(s)d(s) + yisMs) = 1 with
deg y(s) < deg d(s)
[or deg x(s) < deg n(s)].
(2.26) (2.27)
Proof. In view of Theorem 2.4, d(s) and n(s) are coprime if and only if there exist polynomials x(s) and y(s) such that x(s)d(s) + y(s)n(s) = 1. The basic Division Theorem for polynomials can now be used to obtain polynomial solutions of lower degree. In particular, there exists a unique quotient polynomial q(s) and a unique remainder polynomial
541
r{s) such that y(s) = q(s)d(s) + r(s)
with deg r(s) < deg d(s).
CHAPTER 7 :
Now define y(s) = r(s) = y(s) - q(s)d(s),
(2.28)
x(s) = x(s) + q(s)n(s),
where degy(s) < degd{s). Then x{s)d{s) + y{s)n{s) = x(s)d(s) + y(s)n(s) + [q(s)n(s)d(s)-q(s)d(s)n(s)] = x(s)d(s) + y(s)n(s) = I. Also note that deg [x(s)d(s)] = deg [I - y(s)n(s)] < deg[d(s)n(s)], so that deg x(s) < degn(s). To prove uniqueness, suppose there exists another pair x(sX y(s) that satisfies the lemma. Then [x{s) - x{s)]dis) + [>^(^) - y(s)]n(s) = 0, which imphes that [xW-xW] =
-[y(s)
d{s)
Since deg [y{s) - y{s)\ < deg d(s) and the left-hand side is a polynomial, this relation implies that d(s) and n(s) must have a common factor that contradicts the assumption that d(s) and n(s) are coprime. Therefore, x(s) and y(s) are unique. Note that the proof of the lemma when in (2.27) deg x(s) < deg n(s) (for the part in parentheses) is completely analogous. In this case, let x(s)d(s) + y(s)n(s) = 1 and divide x(s) by n(s). • Given coprime polynomials d(s) and n(s), one way of determining x(s) and y(s) that satisfy the above lemma is to follow the procedure outlined in its proof. In particular, x(s) and y(s) that satisfy (2.25) are determined first, and then polynomial division is used to reduce their degrees, if necessary. The following example illustrates this procedure. EXAMPLE 2.19. Let d(s) = s(s - 1) and n(s) = s 2, which are coprime. Let = 1. x(s)d(s) + y(s)n(s) = \s[s(s-l)]-l(s^ + s + 2)[s-2] = k(s'- ^)-Us^-s^-4) Then y{s) • ^Us^ + s + 2) = q(s)d(s) + r(s) = (-\)[s(s - 1)] + [-\{s + 1)], from which we obtain yi^s) = r{s) = -\{s+
1)
and
x{s) = x(s) + q(s)n(s) = ^s + (-\)(s
- 2) = ^
Note that x(s)d(s) + y(s)n(s) = ^[s(s - 1)] + (-\(s + l))ls - 2] = 1 and degy(s) = 1 < degd{s) = 2 [degx(s) = 0< degn(s) = 1]. The polynomials x(s) and y(s) can also be determined using the Sylvester Matrix of the polynomials d(s) and n(s), which in fact provides an alternative way of testing the coprimeness of d(s) and n(s). The Sylvester Matrix of d(s) and n(s) is now introduced. Consider the polynomials, d(s) = dnS"" + dn-lS''^^
+ '" + diS + do,
dn^O
n(s) — riynS^ + nyn-\s^~^ + • • • + n\s -\- n^,
m < n.
The Sylvester Matrix of the polynomials d{s), n(s) is defined as dl do 0 0 dn dn-l dn-2 '" 0 0 •• d2 dx do dn dn-l
S(d, n) =
0
0
^m
rim-i
0
^m
0
0
0 nm-2 nm-\
•• 0 •• 0
" 0 • • no ' ' •' nx
dn
dn-l
dn-2
' •• do
0
0 0
0 0
•• 0 •• 0
0
0
no
nm-i
(2.29)
no
(2.30)
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
542 Linear Systems
where the block that contains the coefficients of d(s) has m rows and the one that contains the coefficients of n(s) has n rows. The coefficients are arranged so that there art n + m c o l u m n s , i.e., S(d, n) E Rin+m)X{n+m)^
The determinant of S{d, n) is known as the Sylvester Resultant or Resultant of the polynomials d{s) and n{s). Note that the matrix S{dy n) is sometimes^eferre4ta^s the Eliminant Matrix of the polynomials d{s), n(s). • THEOREM2.7. The polynomials d(s) and n(s) are coprime if and only if det S{d, n) 7^ 0, that is, if and only if their resultant is nonzero. Proof, (Necessity) Suppose that d{s) and n{s) are coprime but their resultant is zero. This implies that there exists a nonzero vector a, a^ G /?""+" such that aS{d, n) = OOY such that
's'^-^disY ^n+m-l
sd(s) d(s) = a = s^'-^nis)
^n+m-2
aS{d, n)
[ai(s),a2(s)]
d(s) = 0, n(s)^
S
1 J
where 0:1(5) = a i
sn(s) . n(s) J
anda2W = «2
with [01,0:2] = a. Note that both oiC^y)
and 0:2(5) are nonzero since otherwise either n(s) or d(s) would be zero and therefore n(s) and d(s) would not be coprime. Write d(s)/n(s) = -a2(s)/ai(s). Since deg02(5) < degd(s) and degai(s) < degn(s), it follows that d(s) and n(s) must have a common factor. Therefore, d(s) and n(s) are not coprime, which is a contradiction. Thus, S(d, n) has full rank, or detS(d, n) # 0. (Sufficiency) Assume that detS{d, n) ^ 0. Take a = [ 0 , . . . , 0, l]S(d, n)'^ and let 0:1(5) = o:i[5'"~^ . . . , 5, 1]^, 02(5) = 02[5'^~^ .. .,5, 1]^, where [ai, 02] = a. Then 0:1(5)^(5) + 0:2(5)^^) = o:5(J,^)[5"+^-i,5"+'^-2, . . . , 5 , 1]^ = 1, which implies, in view of Theorem 2.4, that d(s) and n(s) are coprime. • Remarks (i) Theorem 2.7 is attributed to Sylvester (1840). (ii) A useful relation, used in the proof of Theorem 2.7, is
s'^-^dis)] ^n+m-\ ^n+m—2
S(d, n)
sd{s) d(s)
(2.31)
sn(s) n(s) where d(s) and n(s) are given in (2.29) and S(d, n) is defined in (2.30).
(iii) It is possible to arrange the coefficients in S{d, n) in several different ways, recalling that elementary row and column operations leave the rank of S{d, n) unchanged. For example, forn = 3, m = 2, ^3
0 (a)
«2
0 0
di d3 n\ «2
0
d\ d2 no n\ ni
0"
do d\
do
0
0 ,
no
0
n\
no J
(c)
do 0 no 0
di do ni no
0
0
d^
^2
0
d3
(b) 0
0
0 _n2
n2 ni
d2 di n2 ni no
dx d2 n2 ni no
do d\ ni no
0
0 do no
0 0
d3 0" d2 d3 0 0 n2 0 n 1 n2
are all nonsingular if and only if d(s) and n(s) are coprime. The matrix in (a) is the Sylvester Matrix as defined in (2.30). The matrices in (b) and (c) are variations of the Sylvester Matrix found in the literature, (iv) By adding zero coefficients in n(s), it can always be assumed that the polynomials d(s) and n(s) have equal degrees (m = n) when using the test provided by Theorem 2.7. This leads yet to another variation of the Sylvester Matrix of two polynomials of dimension 2n X 2n. Using the procedure followed in the sufficiency proof of Theorem 2.7, it is straightforward to derive the unique polynomials x(s) and y(s) that satisfy (2.26) and (2.27) of Lemma 2.6. This is illustrated in the next example. EXAMPLE2.20. Lti d(s) = s(s-l) midn(s) = ^ - 2 , which are coprime as in Example 2.19. In this case the Sylvester Matrix or the eliminant of d(s) and n(s) in (2.30) is n -1 01 S(d, n) = \l -2 0 and the resultant is detS{d, n) = 2 T^ 0, as expected, since LO 1 -2J d{s) and n{s) are coprime polynomials. In view of the sufficiency proof of Theorem 2.7, [4 - 2 0 let a = [ai,a2\ = [0,0,l]S(d, n)'^ = [0,0,1] [ i i,-ij.Then x(s) = ai(s) = ^,y(s) = a2(s)
- ^(5 +1), which also is the answer
determined in Example 2.19. Using the Sylvester Matrix of coprime polynomials enables one in fact to determine solutions to the more general Diophantine Equation x(s)d(s) + y(s)n(s)
q(sl
(2.32)
as the following lemma shows. LEMMA 2.8. Let d(s) and n(s) be coprime polynomials with degd{s) = n and degn{s) = m ^ n given by (2.29). Let q(s) be a polynomial of degree n + m - 1. Then there exist unique polynomials x{s) and y{s) that satisfy x{s)d{s) + y{s)n{s) = q(s) with
f x(s) < m
and
deg y(s) < n.
(2.33) (2.34)
543 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
544 Linear Systems
Proof, The proof is constructive. Let x{s) = [x^-i,. • •, xi, xo][s'^ ^,.. .,s,l]^, let y(s) = [yn-h-",yhyo][s''~K...,s,lV,2indwntQx(s)d(s) + y(s)n(s) = [ x ^ - i , . . . , XQ, yn-h . •., yo][s'^~^d(s),..., sd(s), d{s), s^'-^nis),..., sn{s), n{s)Y = [Xm-h • • •, -^o, yn-u . . . , yo]S(d, n) [s-^^-\ . . . , ^, 1]^ = q(s) = [qn^^-i, •.., ^i, q^Ms^^^'K •. •, ^, 1 ] ^ where relation (2.31) was used. This equality is now satisfied if and only if [Xm-\, ...,XQ,yn-\,...,
yQ\S{d, u) - [qn+m-\, - ", q\,
(2.35)
is satisfied. Since S{d, n) is nonsingular, Eq. (2.35) has a unique solution that determines unique x{s) and y{s) that satisfy relations (2.33) and (2.34). • EXAMPLE 2.21. Consider d{s) = s{s - 1) and n{s) = s - 2, which are coprime as in Examples 2.19 and 2.20. Let q(s) = q2S^ + qis + qo be an arbitrary seconddegree polynomial (n + m - I = 2 + 1 - 1 =2). Relation (2.35) yields in this
"1 - 1 0" case [xo,yi,yo] 1 - 2 0 = [q2, qi, qo]^ from which we obtain [XQ, yi,yo] = .0 1 -2. '4 - 2 0' 0 = [2q2 + ^1 + ^0, -qi q\ - hqo, -\qo\ and \[q2,q\,qo\ 2 - 2 .1 - 1 - 1 , x{s) = XQ = 2q2 + q\ y(s) = [yi, yo] For q{s)
\qo
= (-\qo)
+ {-qi - q\-
\qQ)s.
+ 2 ^ + 1 , x{s) = | , and y(s) = - \ - Is.
•
COROLLARY 2.9. Let d(s) and n(s) be coprime polynomials with degd(s) = n and deg n(s) < n. Let q(s) be a polynomial with deg q(s) < 2n- I. Then there exist unique polynomials x(s) and y(s) that satisfy the relation x(s)d(s) + y(s)n(s) = q{s) with
deg x{s) < n
and
deg y{s) < n.
(2.36) (2.37)
Proof, The proof is similar to the proof of Lemma 2.8, where in place of S{d, n) E ^(n+m)x(n+m)^ the 2n X 2n Sylvester Matrix S{d, n) of d{s) = dnS"" + • • • + Ji^ + JQ, (i„ 7^ 0 and n{s) = rinS^ + --- + nis + HQ is used. Zero coefficients are added in n(s) if necessary. Note that S(d, n) is exactly S(d, n) of (2.30) with m = n. This leads to the relation {Xn-\, ...,XQ,
yn-\,...,
y()\S{d, n) = [qm-i, • • •, ^b ^oL
(2.38)
which is the equation corresponding to (2.35). Since d{s) and n{s) are coprime, S{d, n) is nonsingular and the x{s) and y{s) that satisfy (2.38) are unique. • EXAMPLE 2.22. As in Example 2.21, consider d{s) = s{s — \) = s^ - s and n{s) = 2 = O'S^ -\-s- 2, which are coprime. Then (2.38) implies that
Ti 0 [xi, xo, yi, yo] p 0
-1
0
1 -1 1-2 0 1
ol 0 = [qs, qi. qi, qol 0 -2
where ^(5-) = ^3^^ + ^2-y^ + ^i^ + ^o is a third-degree polynomial ( 2 ^ - 1 = 2 . 2 - 1 = 3). It can easily be seen that one could equivalently solve xi = qs and [XQ, yi, yo] X "1 - 1 01 0 = [qi + q-iy qu qo] (compare with Example 2.21), from which we obtain -2
x(s) = [xi, xo]
545
= qss + [2(q2 + ^3) + ^1 + ^qo]
CHAPTER?:
= [-(qi + qs) - qi-
y(s) = [yu yo]
\qQ\s + (-5^0).
Note that for q^ = 0, this solution is precisely the solution of Example 2.21.
•
The polynomial matrix case Similar results as in the preceding can be shown for polynomial matrices. First the Division Theorem for polynomial matrices is established. THEOREM 2.10. Let D{s) E R[sT^'^ be nonsingular. Then for any A^(^) G RisY^"^ there exist unique polynomial matrices Q{s), R(s) such that N(s) = Q(s)D(s) + R(s)
(2.39)
with R(s)D(s) ^ strictly proper, or with deg^. R(s) < deg^. D(s), j = 1 , . . . , m, when D(s) is column reduced. Proof. Let H(s) = N(s)D~^(s) be a rational matrix. By polynomial division in each entry, decompose this matrix uniquely into H(s) = Hsp(s) + P(s), where Hsp(s) is a strictly proper rational matrix and P(s) is a polynomial matrix. Then N(s) = H(s)D(s) = P(s)D(s) + R(s), where R(s) = Hsp(s)D(s). Then R(s) is a polynomial matrix (since it is equal to N(s) - P(s)D(s)), and by definition, R(s)D~^(s) = Hsp(s) is strictly proper. Note that when D(s) is column reduced, then in view of Lemma 2.2, R(s)D~^(s) is strictly proper if and only if deg^. R(s) < deg^. D(s), j = 1 , . . . , m. We shall now verify uniqueness of Q(s) and R(s) directly; it also follows from the construction. Suppose there are two pairs such that Q(s)D(s) + R(s) = Q(s)D(s) + R(s). Then Q(s) ~ Q(s) = [R(s) — R(s)]D~^(s). Since the left-hand side is a polynomial matrix and the right-hand side is a strictly proper rational or zero matrix, this equality can hold only when both sides are zero, i.e., Q(s) = Q(s) and R(s) = R(s). • As expected, the following result also holds. THEOREM 2.11. Let D(s) G R[sY^P be nonsingular. Then for any N(s) G RlsY^"" there exist unique polynomial matrices Q(s), R(s) such that N(s) = D(s)Q{s) + R(s)
(2.40)
with D~^(s)R(s) strictly proper, or with deg^. R(s) < deg^. D(s), i = 1,..., p, when D(s) is row reduced. Proof. The proof of this result is completely analogous to the proof of Theorem 2.10. • Given re polynomial matrices D(s) G R[s]^^^ and A^(^) G R[sy^"^ it was shown in Theorem 2.4 that there exist polynomial matrices X(s) and Y(s) such that X(s)D(s)
+ Y(s)N(s)
= Im.
(2.41)
Let H(s) = N(s)D-\s)
=
D-\s)N(s)
(2.42)
be proper, where D(s) G R[sy^P and N(s) G R[sy^^ are Ic with D(s) row reduced. Let V be the highest degree of the polynomial entries in D(s) and denote this as p =
degD(s).
(2.43)
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
546
We point out that it can be shown that v is the observabihty index of the system
Linear Systems
His) = N{s)D-\s)
= D-\s)N{s).
LEMMA 2.12. Given D{s) G RisT^"^ and N{s) E RisY^"^ that are coprime, there exist polynomial matrices X{s) and Y{s) that satisfy (2.44)
X{s)D{s) + Y{s)N{s) such that deg Y(s) < V.
Proof, This result can also be established by using the eliminant of N{s) and D{s) to be introduced in Lemma 2.14. In the present proof, use is made of the Division Theorem for polynomial matrices (Theorem 2.10). Let X{s) E Risf'''^ and Y{s) E R[SY'''P satisfy (2.41). There exist unique polynomial matrices Q{s) E R[ST^P and R{s) E RisT'^P such that Y{s) = Q{s)D{s) + R{s) with R{s)D'^{s) strictly proper, where b{s) is row reduced and satisfies (2.42). Now define Y{s) = R(s) = Y(s) - Q(s)D(s) and X(s) = X(s) + Q(s)N(s). Then X(s)D(s) + Y(s)N(s) = X{s)D(s) + Y(s)N(s) + [Q(s)N(s)D(s) Q(s)D(s)N(s)] = X(s)D(s) + Y(s)N(s) = /^. Note also that, in view of Lemma 2.1, if R(s)D'^(s) is strictly proper, then the column degrees of R(s) are less than v, i.e., deg Y(s) < V. In addition, note that from X{s)D{s) = Im - Y(s)N(s) it follows that deg,. (X(s)D(s)) = deg,. (/^ - Y(s)N(s)) I, j = 1 , . . . , m. The Eliminant Matrix Me of N(s) and D(s) is a generalization of the Sylvester Matrix, or of the Eliminant of two polynomials, introduced in (2.30), and is defined in an analogous manner (see also Remarks following Theorem 2.7). In particular, we write A^(^) sN(s)
D(s) sD(s) e^-l
D(s)
= MekSek(s)
= Mek
block diag
(2.45) ^dj + k-\
for some integer k, where M^k G RKp+rn)x{j^dj+mk)^ ^^^^ ^^^^ f^^ ^ > {ldj)/p,Mek will have as many as or more rows than columns. Let ^ = V be the least integer k that minimizes (^dj + mk) — rank M^k, or equivalently (as can be shown), let ^ = V denote the observability index of the system {DJ^N}, or of its equivalent state-space description (refer to the discussion on equivalence of representations in Subsection 7.3A). The Eliminant Matrix of N{s) and D{s) is defined as the v{p + m) x{n + mv) matrix Me=Mev, where n = ^dj = deg det D{s), since D{s) is column proper. Note that the observability index v satisfies v > n/p, and therefore, Me has as many as or more rows than columns. The following result is a generalization of Theorem 2.7 (which involved the Sylvester Matrix of two polynomials) to polynomial matrices. THEOREM 2.13. The polynomial matrices N{s) and D{s) given above are re if and only if their eliminant matrix M^ has full column rank, i.e., if and only if rank Me = n + mv.
(2.46)
Proof, The proof is omitted. For a proof of this theorem, refer to Wolovich [36]. Remarks (i) The Eliminant Matrix Me becomes polynomials. In particular, for d{s) n^s^ -\-n2S^ -\-nis-\-no^p = m= l,/i n{s) sn{s) d{s) sd{s) s^d{s)\ (ii)
•MeSeis)
"no 0 0 do 0 .0
•
a Sylvester Matrix when applied to = ^3^^ + d2S^ -\-dis-\-do and n{s) = di=v = 3, this is true where
n\ n2 m 0 0" no ni ni m 0 0 no n\ ni "3 d3 0 0 di di do di di d3 0 0 do di di d3.
1
See also Remarks following Theorem 2.7. It can be shown that rankMek = rankMey,
^ > V,
i.e., the rank of Mek is the maximum possible when k is equal to the observability index of the system. This is a consequence of the alternative definition for v previously given. The Eliminant Matrix can be used to determine solutions of the Diophantine Equation X{s)D{s)^Y{s)N{s) = Q{s). (2.47) The Diophantine Equation is further discussed later in this subsection. LEMMA 2.14. Consider the re polynomial matrices N(s) e R[S]P'''^,D(S) e R[s]'^'''^. Let D{s) be column proper, and assume deg^.N{s) < deg^.D{s) = dj,dj > lj = 1,...,m [note that H{s) = N{s)D-^ (s) is proper]. Let Q{s) G Rls]^"""^ with deg,.Q{s)
(2.48)
547 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
548 Linear Systems
where v is the observabiHty index of the system {D, I, N, O}. Then there exist X(s) R[s]qxm and Y(s) G R[s]^''P that satisfy the equation X(s)D(s) + Y(s)N(s) = Q(s) with
degX(s) < v ~ I,
(2.49)
degY(s) < v - I,
where degX(s) denotes the highest polynomial degree in the entries of X(s). Proof. The proof is by construction. Let In. Sim
X(s) = Xo+Xis
+ '-'+X^-is""^
= [Xo,Xu
(2.50)
...,X,-i] 3
Iffi
sin
and
(2.51)
Y(s) = Fo + Fi5 + • • • + F.-i^""^ = [YQ, FI, . . . , F,-i] S^'-^In
N(s) sN(s)
Then Y(s)N(s) + X(s)D(s) = [YQ, F ^
Yp-i,Xo,Xi,..
.,Xj,-\]
D(s) sD(s)
[Fo,..., F^-i, Z o , . . . , X^,-i]MeSe(s) = Q(s) = QSe(s), in view of the definition of the eliminant matrix Me and the assumptions on the column degrees of Q(s) in (2.48). Therefore, (2.49) is satisfied with X(s) and Y(s) given in (2.50) and (2.51) if and only if [Fo,...,F,_i,Xo,...,X,_i]M, = Q
(2.52)
is satisfied, where M, G R^(p+m)x(n+mv) ^^^ Q ^ j^qx(n+mv) ^^^^ ^ ^ ^"J^i^j = degdetD(s). Since N(s) and D(s) are re, it follows from Theorem 2.13 that rank Me = n + mv, i.e., Me has full column rank. This implies that a solution to (2.52) exists for arbitrary Q. Therefore, solutions X(s) and Y{s) of (2.49) with degX(s) ^ v - I and deg Y(s) < z^ - 1 always exist. • Remark When N(s) and D(s) are polynomials, Lemma 2.14 reduces to Corollary 2.9 (with n(^)/(i(5*) proper). In this case, v = di = n = deg(detD(s)). EXAMPLE 2.23. The polynomial matrices N(s) = are re since rank
rA^(A)
LD(A).
s+l 1
0 ,D(s) 1
0 -s + 1
2 for A = 0 and A = 1, the roots of detD(s). D(s) is col-
umn proper with deg^ D(s) = d\ = 2, deg D(s) = di = \, n = d\ -\- d2 = 3, while deg^^ N(s) = 1 < Ji = 2 and deg^^ N(s) = 0 < d2 = I, p = m = 2. To construct the Eliminant Matrix of N(s) and D(s), we note that n/p = | , and therefore, we let A: = 2
be an initial value. Then
549 1 1
1 0
0 0
0 0
0 1
0 0
CHAPTER?:
ri N(s) sN(s) D(s) sD{s)
0 0
1 1
1 0
0 0
0 0
0 1
0 1
0 0
1 0
0 0
0 1
0 -1
0 0
0 0
0 1
0 -1
0
01 0 0 0 1 s
[o
s^\
0 s 0 \s^
r'0
MelSe:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
Note thai rank Me2 = 1 = ^z + m^, which is full column rank. Therefore, z^ = 2 and the EHminant Matrix of N(s) and D(s) is Me = Mei- The fact that rank Me = 1 also verifies that N{s) and D{s) are re. We consider the Diophantine Equation (2.49) next, (i) First, we let Q{s) = [s^ + 1, s^], which satisfies (2.48), and we write Q(s) = I s s^ s^ 0 0 0 QSe(s) = [1, 0, 0, 1, 0, 0, 1] 0 0 0 0 1 5 ^2 . Then Eq. (2.52) assumes the form [Yo, Yi,Xo,Xi]Me
= Q = [l 0, 0, 1, 0, 0, 1],
which is an algebraic system of linear equations with 8 unknowns and 7 linearly independent equations that have more than one solution. One such solution is given by [70,1^1,^0,^1] = [I 0 , - 1 , 1, 1, 0, 1, - 1 ] , which, in view of (2.50) and (2.51), implies that X(s) = Xo + Xis = [1, 0] + [1, -l]s
= [s+l,
-s]
Y(s) - Fo + Yis = [1, 0] + [ - 1 , 1]^ = [1 - s,s] is a solution to the Diophantine Equation with degX(s) degY(s) = \ = v - \ . (ii) Next, we let Q(s) = ^ 1 0
0
0 : 0 0 0
0 0 0 0 : 1 0 0 assumes the form [Yo,YuXo,Xi]Me
= I = v
1 and
and we write Q{s) = QSeis) = 1 0
s 0
0
0 1
0 s
1 0 0 0
0 0
0 0
0
= Q =
A solution of this equation is [FQ, FI, XQ, Xi ] =
Then Eq. (2.52)
0 0 1 0
0 0
1 0 - 1 0 1 0 0 0 - 1 1 1 0 - 1 0 0 0
This implies that X(s) = Xo+ Xis Y(s) = Yo + Yis
10] ro 0 -1 oJ"^Lo 0 1 01 r - 1 0' -1
1
+
1
0
1 -1 1 -s - 1 +s
0 0 0 1
is a solution of the Diophantine Equation with deg X(s) = 0 < z ^ - l = 1 and
degY(s) = \ = v-\.
•
550 Linear Systems
In the preceding development, it was shown how to derive particular solutions of the Diophantine Equation, given coprime polynomials and polynomial matrices. In the following, conditions for existence of solutions are derived and all solutions of the Diophantine Equation are conveniently parameterized. THEOREM 2.15. Consider nonzero polynomial matrices D(s) G RlsY^^"^ and N(s) G R[s]P2Xm with pi+ p2^ m, let GR(S) G /?[^]^^^ be a gcrd of D(s) and A^(^), and let Q(s) G R[sY'''^. The Diophantine Equation (2.53)
X(s)D(s) + Y(s)N(s) = Q(s)
has polynomial matrix solutions X(s) G R[SY^P^ and Y(s) G R[S]^^P^ if and only if GR(S) is an rd of Q(s). If GR(S) is an rd of Q(s), then (2.53) has infinitely many polynomial matrix solutions. If X(^) = Xo(s) and Y(s) = FQ(5) is one such solution, then all solutions of (2.53) are given by X(s) = Xo(s) -
K(s)Nds)
(2.54a)
Y(s) = Yo(s) +
K(s)Ddsl
(2.54b)
where K(s) G Ris]^""' and where NL(S) G RIS^^P^, are Ic and satisfy NL(S)D(S)
of the left kernel of
+ DL(S)N(S)
DL G RISY^'P^, r = (pi + P2) - m
= 0, i.e., [-NL(S),
DL(S)] is a minimal basis
N(s),
Proof, If there exist X and Y that satisfy (2.53), then (XDR + YNR)GR = Q, which implies that QG^ ^ must be a polynomial matrix. Thus, GR is an rd of Q that is a necessary condition for solutions of (2.53) to exist. To show that this also constitutes a sufficient condition, assume that GR is an rd of Q, i.e., Q = QRGR for some polynomial matrix QR, and recall that there exist polynomial matrices X G RisY'^P^ Y G RISY'^P^ such that (2.55)
XD + YN
The proof of this result is constructive and is based on the algorithm to determine a gcrd of D and N, using the Euclidean algorithm. (Refer to Subsection 7.2D, where it is shown that GR X Y , where [/ is a unimodular matrix.) Premultiplying U 0 -NL a or (2.55) by QR, we now obtain (QRX)D + (QRY)N = QRGR XQD
+ YoN = Q,
(2.56)
i.e., GR being an rd of Q is also a sufficient condition for the existence of solutions of (2.53). If (2.56) is satisfied, then (Xo - KNL)D
+ (Fo + KbL)N
= XoD + YoN = Q,
and therefore, (2.53) has infinitely many solutions, among them the infinitely many given by (2.54). It remains to be shown that every solution of (2.53) can be put into the form (2.54). Suppose that X and Y satisfy (2.53). Subtracting (2.56) from (2.53), we obtain (X - Xo)D + (Y- Yo)N = 0, or [X - Xo, F - Fo] LQiNi^RisY''^'
and^L
D
RlsY^'P^, r = (pi
0. m be relatively Ic and such that
= 0.
Then [-NL, DL] is a minimum basis of the left kernel of
. (A discussion of polynomial
bases of vector spaces will be given shortly.) This implies that there exists some K E Rls]^""' so that [X-Xo.Y-Yo]
=
K[-NL,DLI
which is (2.54). Thus, every solution of (2.53) can be written in the form of (2.54).
•
Remarks (i) If Z)-i exists, thenin the proof of Theorem2.15,(X-Xo)D + (F-7o)A^ = 0 implies that Z - Z o = -(Y-Yo)ND-^ = - ( F - F o ) ^ Z ^ ^ z . , where A^L and DL are Ic. Since X - XQISSL polynomial matrix, this implies that Y - YQ = i'^DL for some iT, where X-Xo = -^A^L-Thus, when i)~^ exists, the above results can be derived without directly using the concept of minimum basis \D] of the left kernel of A^ (ii) In the theory of systems and control, the Diophantine Equation is of great use, particularly for the case when D~^ exists and ND~^ is a proper rational matrix. In this case solutions of the Diophantine Equation with special properties are of interest, particularly solutions where X~^ exists a n d Z ' ^ F is a proper rational matrix (see Subsections 7.4A, 7.4B, and 7.4C). Such solutions, satisfying particular properties, may also be determined via polynomial matrix interpolation (see the Appendix for theory and examples). It will be shown later in this chapter that the Diophantine Equation is central in the study of systems wherever feedback is used. Rings, modules, and polynomial bases of rational vector spaces Now some useful concepts from algebra are briefly reviewed. In particular, rings of polynomials, of proper rational functions, and of proper and stable rational functions are briefly discussed, together with modules and polynomial bases of rational vector spaces. Consider the set of polynomials R[s] in s with real coefficients, together with the usual operations of addition, +, and multiplication, •, of polynomials. Then (R[s], +, •) is a ring, since all the axioms of a ring are satisfied. These axioms are the same as the axioms of a field (see Section 1.10 of Chapter 1), except for axiom (vii), which assumes existence of the multiplicative inverse in a field. Indeed, if p(s) G R[s] then l/p(s) ^ R[s], since in general l/p(s) is a rational function but not a polynomial. Another example of a ring is the set of integers, taken together with the usual operations of addition and multiplication on integers. We note that R[s] is a Euclidean ring, i.e., (i) for a{s), b(s) G R[s] such that a(s)b(s) 7^ 0,deg[a(s)b(s)] > dega{s)\ and (ii) for every a{s),b{s) G R[s^ with b{s) ¥- 0, there exist ^(5-), r{s) G R\s\ such that a(s) = b(s)q(s) + r(s), where either r(s) = 0 or deg r(s) < deg b(s) (polynomial division). If R{s) is the field of rational fractions over R[s], called the field of rational functions (see Subsection 1.2A), then R[s] C R{s) and R{s) is also a Euclidean ring. Note that the units of R{s] are those polynomials p{s) for which there exist a p\s) G R[s] such that p(s)p'(s) = 1, the unity element of R[s], i.e., the units of R[s] are the nonzero elements of R. Another Euclidean ring of interest is the set of proper rational functions, taken together with the operations of addition and multiplication. A proper rational function
551 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
552 Linear Systems
t(s) = n(s)/d(s) (d(s) # 0) is a unit of this ring if deg n(s) = deg d(s), i.e., if it is ^ biproper rational function. The set of proper and stable rational functions (with poles in the stable region of the ^--plane) is also a Euclidean ring. In this case, if t{s) is a unit then t{s) and t~^{s) are proper and stable rational functions. Polynomial bases of rational vector spaces are discussed next. Recall that if Pi{s) G R[s], then it may always be viewed as being divided by 1, in which case Pi{s) = Pi(s)/1 E R(s), the field of rational functions. Thus, a polynomial vector p(s) E R[s]P may be viewed as a special case of a rational vector p(s) E R(sy. Now let H(s) E R(sy^^, assume that rankH(s) = m, and consider the rational vector space spanned by the columns hj(s) E R(sy, j = 1 , . . . , m, of H(s) and denote this space by Y(s). Thus, the vectors in Y(s) are generated by H(s)a{s), where a(s) = [ai(s),..., ap(s)]^, cxj(s) E R(s). The matrix H(s) is a basis for Y(s) and if T(s) E Ris)"^"""^ and rankT(s) = m, then H(s) = H(s)T(s) is also a basis for Y(s). Consider a right polynomial MFD of H(sl H(s) = B(s)A~\s), where B(s) E R[sy^^ and A(s) E R[s]^^^ are not necessarily coprime. Note that rankB(s) = m and B(s) is a polynomial basis of the rational vector space Y(s). Consider now the set M of all polynomial vectors h(s) E R[sy that can be written as linear combinations over the ring R[s] of the columns bj(s) E R[S]P, j = 1 , . . . , m, of B(s). The set M is 3. free R[s]-module and 5(^) is a basis of M. The dimension of M is
problems, is considered to be one of the greatest mathematicians. His treatise on algebra, considered to be the first ever, is called "API0MHTIKA" and consists of seven books, six of which have survived. Diophantus worked with equations involving integers. An example of a specific case of the Diophantine Equation is 3x-\-2y = 4, the general solution of which is given b y x = 2-2k,y = - 1 + 3^ with k equal to any integer (see Theorem 2.15). In this case, the particular solution XQ = 2, yo = -I was used. The parameterization of all solutions in this convenient form appears to be a later development, the work of Hindu mathematicians in the fifth century A.D., apparently of Aryabhata. It is worth mentioning at this point that another class of famous Diophantine Equations is x^ + y^ = z^, where x, y, z, and n are integers. When n = 2, then x = 3,y = 4, and z == 5 is a solution since 3^ + 4^ = 5^. Solutions for integer n greater than 2 do not appear to exist. In fact, the famous Fermafs last theorem (c. 1637) states that no solution exists for ^ > 3. Fermat wrote on the margin of his copy of the works of Diophantus, referring to this theorem, "I have discovered a truly remarkable proof which this margin is too small to contain." In spite of attempts over the next centuries by mathematicians like Euler, Legendre, and many others, this result has not yet been proved in its generality, although a recent proof (1994) shows great promise of being completely correct. It is, however, known to be true for n up to tens of thousands, using computer methods. Diophantus worked with integers. However, integers and polynomials obey similar rules (they are both rings), and therefore, all results of interest on the linear Diophantine Equation, xd -\- yn = q that were developed for integers can readily be written for polynomials, and this is what was done in Subsection 7.2E. These results for the polynomial Diophantine Equation were pointed out in the books by Gantmacher and McDuffee (refer to the bibliography at the end of the chapter). In view of this, it is not surprising that solutions of linear Diophantine Equations involving elements other than integers or polynomials, but still members of rings, can be studied and expressed in a completely analogous manner. In fact, Diophantine Equations with elements that are polynomials in z~^ and (s -f- a)~^ were studied in the systems and control literature in the 1970s by Kucera and Pernebo, among others. Later, in 1980, Desoer et al. clearly pointed this fact out in the systems and control literature, namely, that the solutions of the linear Diophantine Equation, when the equation involves elements of rings other than integers or polynomials, are analogous to the integer and polynomial cases. They also addressed conditions on system descriptions under which such Diophantine Equations may appear. It will be shown in Subsection 7.3C how the linear Diophantine Equation is of fundamental importance in the study of feedback systems. It should be noted here that when q = 1, the Diophantine Equation is sometimes referred to as the Bezout identity (see, e.g., Kailath [21]). For the role played by the Diophantine Equation in control, the reader should also refer to the survey paper by Kucera [24]. For further details, refer to Sections 7.6 and 7.7, Notes and References, respectively. 7.3 SYSTEMS REPRESENTED BY POLYNOMIAL MATRIX DESCRIPTIONS We consider system representations of the form P(q)z(t) = Q(q)u(tl
yit) = R(q)z(t) +
W(qMt)
(3.1)
553 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
554
Linear Systems
with P{q) G R[q]^''\ Q{q) G R[qY'''^, R{q) G R[q]P''K and W{q) G R[q]P'''^, wliere R[qy^^ denotes tlie set of / x / matrices whose entries are polynomials in q with real coefficients {q = (d/dt) the differential operator) and we assume that det P{q) ^ 0 and that u{t) is sufficiently differentiable. We also assume that the set of homogeneous differential equations P{q)z{t) = 0 is "well formed," that is, for all initial values of z(-) and its derivatives at f = 0 the solution does not contain impulsive behavior at f = 0. (This is true if and only if P~^{q) is a proper rational matrix, which is true for example when P{q) is column or row reduced (for further details refer, e.g., to Section 3.3 of [9])). Such representations, called system Polynomial Matrix Descriptions (PMD), were introduced in Section 7.1. In this section, equivalence of system representations is discussed first, in Subsection A. This notion of equivalence establishes relations not only between polynomial representations, but also with state-space representations, which are in fact a special case of PMDs (see Section 7.1). Equivalence of PMDs preserves certain system properties, including controllability, observability, and stability, that are the topics of discussion in Subsection B, together with polynomial matrix realizations of transfer function matrices. In Subsection C, Polynomial Matrix Fractional Descriptions (PMFDs) are employed to study properties of interconnected systems.
A. Equivalence of System Representations Consider the PMDs Pi{q)zi{t) = Qi{q)u{t), y{t)=Ri{q)zi{t)+Wi{q)u{t),
1,2,
w\\QXQPi{q)eR[q]^'''^',Qi{q)eR[qy''''^,Ri{q)eR[q]P''^\
(3.2)
and W,-(^) G/^[^]^>^
DEFINITION 3.1. The representations {Pi, gi,/?i,\^i} and {P2,22,^2,^2} in (3.2) are called equivalent if there exist polynomial matrices M ^ R[q\^^^^^ ,N ^ R[q\^^^^^ ,X ^ RW^K and Y G P[^]^2xm ^^^^ ^^^^ 'M X
0"
h.
Pi
Gi
-^1
Wi
Pi -R2
Qi W2
'N 0
-Y'
(3.3)
^m
with (M, P2) Ic and (Pi, A^) re.
•
The equivalence of the representations {P/,2/,/^/,W^},/ = 1,2, or of the corresponding system matrices ^1,^2, where Pi Qi €R[s] {li+p)^{li+m) -Ri Wi
(3.4)
is sometimes referred to as Fuhrmann's System Equivalence (FSE). This equivalence relation will be denoted by pf, or simply p . It can be shown that (3.3) with its associated conditions given in Definition 3.1 defines an equivalence relation p on the set of matrices S with p, m fixed, \P\ ^ 0, and / any positive integer (see the discussion of equivalence relations at the end of Subsection 7.2C). The polynomial matrix S defined in (3.4) for system {P, g, P, W} is called Rosenbrock's System Matrix of the system Pz = Qu^y = Pz + Wu.
EXAMPLE 3.1. Consider the representations (1.3) and (1.4) in Section 7.1 and let Pi = qI-A,Qi = B,Ri = C,Wx = D given in (1.3), and let P2 = P, Q2 = Q> R2 = 1 0"' ,M = 1 q 0" D R^W2 = W given in (1.4). If A^ = C = [0 0 0 q_ 0 - 1 1, 0 0 andX = 0 0 then (3.3) is satisfied (verify this). Note that (M, P2) = 0 0 0 1. 1 1 ^ 0 ^^ + 1 are Ic, while (Pi, N) are re, as can be verified by applyQ+ 2 0 0 ^ q ing, for example, the rank test for coprimeness of Theorem 2.4(v) to (M, P2) above q 0 1 -1 0 0 '+2 . These representations satisfy all the conditions of and to 0 1 0 0 -1 1 Definition 3.1, and are therefore, equivalent (or Fuhrmann system equivalent).
•
There is an alternative definition of system equivalence, sometimes referred to as strict system equivalence or Rosenbrock's System Equivalence (RSE). Subsequently, PR will denote this equivalence relation. It is defined as follows. The representations given in (3.2) are called strict system equivalent if there exist unimodular polynomial matrices M,N ^ R[q\^^^ and polynomial matrices X G Riq^^, Y G Riqf"^ with m = deg |P/|, k > max(ni, ^2), such that
M
0
0
0
Pi
Gi
Wi h-h
0
0
Pi
Q2
(3.5)
N
0
-R2
:
h, k> h
W2
Relation (3.5) is in general easier to use than (3.3), since the matrices M, A^, and
\N
-Y
M X
0
are all unimodular.
The following theorem establishes that the preceding two definitions of system equivalence [i.e., Fuhrmann's (FSE) (3.3) and Rosenbrock's (RSE) (3.5)] are, in fact equivalent. Note that it was for this reason that in Definition 3.1, which is in fact the definition of FSE, only the term "equivalent" was used. THEOREM 3.1. Consider the representations given in (3.2). If they are Fuhrmann's System Equivalent (FSE) [satisfying (3.3)], i.e., {Pi, Qi, Rx, WI}PF{P2, Qi, Ri, W2}, then they are Rosenbrock's System Equivalent (RSE) [satisfying (3.5)], i.e., {Ph Qh Rh WI}PR{P2, Q2, Ri, W2\ and vice versa.
555 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
556 Linear Systems
-X
Proof, If the representations are FSE, then MPi = P2N and
Pi
I 0
0 , where [-X,Y] and /
Y M
-N Pi
X Y
were chosen so that the block matrices are unimodular.
This can always be done since (P2, M) are Ic and (Pi, N) are re. [See also the discussion on doubly coprime factorizations of a transfer function in (4.18) of Subsection 7.4A.] It can now be seen that -X
Y
P2
M
-R2
X
0
.0
0
Pi
Gi
~Ri
0
Observe that \
0
0
P2
0
-R2
W2
YPi N_
"0
xl \-x ^JU/,
-X
YPi
YQi
h.
N
Y
W
A
(3.6)
which implies that the second block
N.
matrix in the left-hand side is unimodular. Therefore, FSE ^ RSE of the representations. To show the converse, suppose that (3.5) is satisfied. Partitioning, we can write Mil
Mu
M21
M22
Xi
X2
0 0
0
Pi
Wi
-Ri
(3.7)
h-i,
0
0
Nil
N12
-Yi
0
P2
Qi
N21
N22
-Y2
0
-^2
W2 L 0
where M22, N22 ^ ^[^l^^x^i. Then M22
0
Pi -Ri
Pi -Ri
Qi Wi
Qi 'N22 0 W2\
-Y2
(3.8)
Im
It turns out that (M22, P2) are Ic and (Pi, N22) are re. This can be shown as follows. Consider the unimodular matrices M =
Mil
M12
1P2N21
M22.
andA^ =
Nil N21
MnPi N22 J
where
the relations M12P1 = A^i2, M21 = P2N21 from (3.7) were used. Now if (P2, M22) were not Ic, then they would have a nonunimodular common left divisor that would have caused M not to be unimodular. Therefore, they are Ic. Similarly, A^ is unimodular implies that (Pi, N22) is re. This shows that (3.8) is indeed the defining relation for FSE, and therefore RSE => FSE of the representations. •
In the following we enumerate some of the important properties of FSE and RSE. THEOREM 3.2. Assume that the representations (3.2) are equivalent, i.e., they are FSE or RSE. Then the following are invariants of the equivalence relation p = pF = PR(i) deg\Pi\ = nj = 1,2. (ii) The nonunity invariant polynomials (in the Smith form) of P/, [Pi, Qi\ Pi and -Ri Si, i = 1, 2. (iii) The transfer function matrix//,(^) = Ri(s)P;^(s)Qi(s) + Wi(s),i = 1,2. Proof, Recall from Section 7.2 that two polynomial matrices Mi, M2 E R[qy^^ have the same Smith form if and only if there exist unimodular matrices Ui, U2 such that UiMi = M2U2. In this case Mi, M2 are called equivalent polynomial matrices. To show 0 0 (i), note from (3.5) that M N, where M, N are unimodular. 0 0 Pi Pi Therefore, Pi and P2 must have the same nonunity invariant polynomials (note that Pi, P2 may have different dimensions), which implies that |Pi| = a\P2l where a is some nonzero real number (show this). Clearly then, deg \Pi \ = deg IP2I. M 0 To verify (ii), note that in (3.5), are unimodular, and thereand X fore, the extended system matrices 0
h-ii 0
Se, =
^€2
^1
lk-l2
0
0
^2,
~
with Si, i = 1, 2, defined in (3.4), have the same Smith form. Thus, ^1 and ^2 have the same nonunity invariant polynomials. In a similar manner, (3.5) implies that 0
0
Pi
Qi
M 0
0
0
0
Pi
Qi
which shows that [Pi, Qi] and [P2, Q2] have the same nonunity invariant polynomials. Pi have the same nonunity invariant polynomials. Pi and Similarly, -Ri -Ri Finally, to show (iii), we write, in view of (3.5) RiPi'Qi
Hi
+ Wi = [0, Pi] 0 Pi
= [0, R2W + X
=
[0, P2
7
0
0
P2
7
0
0
P2
/ = [0, P2] i 0
0 P2
= [0, P2]
I 0
0 Pi
+ Wi
-1
+ Wi
M +X
M
= [0, P2]
+ Wi
+ 7 0
0
GiJ 0 P2
+ Wi + [W2 + [0P2]?]
+ W2 = RiP2^Qi + W2 = H2.
557 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
558 Linear Systems
In the above derivations the following relations, which can directly be seen from (3.5), were used: M
7 0 0 ^1. M
I 0
7 0" ^ 0 P2_ N,
(3.9)
"0 ' I 0 0 — — f + .Q2_ 0 P2_ Qi.
(3.10)
0 + [0, - / ? i ] = [0, -R2\N, Pi
(3.11)
+ Wi = -{Id,-R2\Y
(3.12)
+W2.
Now consider the representations (3.2) and notice that they can be written as Ziit)
Piiq) -Riiq)
SM) -u(t)
Qiiq) Wiiq)
0
Zi(t)
-uit)
(3.13)
-y(0.
i = 1, 2, where Si(q) is Rosenbrock's System Matrix given in (3.4). If these representations are equivalent, then the relation between the states zi and Z2 can be found in the following manner. Postmultiplying both sides of (3.3) by 1 T -u'Y we obM 0 \ Px Qx Qi tain or = Pi \—u [~Rx Wi [ —M -R2 W2_ M 0 r0 ' 0' ' Pi Qi\ \ Pi Q2] 'Nzi + Yu\' ThereRi W2J \_—u \ — u J X lp\ [-y. [-R 2 W2J -y. fore,
r "^1
\ ^A
Z2(t) = N(q)zi(t) + Y(q)u(t)
(3.14)
is a possible relation between the partial states. To obtain the inverse relation, we could express z\ as a function of Z2 by considering (3.3) and selecting X, Y, X, Y so that
-x .Pi
Y^ \-N M\ [Px
X Y.
I 0
0" /
"-A^
X] \-X
f
Y\ [Pl
M
Pi
(3.15)
are unimodular polynomial matrices. Note that this is always possible, since MP\ = P2N with (M, P2) Ic and {Pi, N) re (see also the proof of Theorem 3.1). Now in view of (3.15), Eq. (3.3) implies that Px -Rx
Qx]\X -E WiJ [0 ^m
x 0] F
Pl
h\ -R2
Qi W2
(3.16)
where £• = YQx-XYandF = i ? 2 ^ - X F . This relation was of course expected, due to the symmetry property of the equivalue relation pp. Postmultiplying both sides of (3.16) by [z\y -u^]^ and proceeding similarly as before, we can now show that Zi(t) = X{q)z2(t) ^ EiqMt).
(3.17)
Equations (3.14) and (3.17) determine an invertible mapping that relates the states of the equivalent representations (3.13). It can be shown that if (3.14) and (3.17) hold for some polynomial matrices A^, Y X, E, then the representations in (3.13) are equivalent.
Polynomial Matrix Fractional Descriptions (PMFDs) Next, we consider the right Polynomial Matrix Fractional Descriptions (rPMFDs) given by Di{q)zi{t) = u{t),
y{t) - Ni{q)Zi{t\
i = 1,2.
(3.18)
THEOREM 3.3. The rPMFDs given in (3.18) are equivalent if and only if there exists a unimodular matrix U such that (3.19)
U. Proof, Suppose that (3.19) is true. Rewrite it as D2
0
(3.20)
0
-N2
and note that this is a relation of the form (3.3). Therefore, the representations {Di, I, Ni}, {D2, /, N2} are equivalent. Conversely, suppose that (3.3) is true, i.e., M X
01 \ ^'
Imi l-Ni
Im
0_
D2 -N2
Im]\^ -Y' oj [0 Im.
with (M, 7)2) Ic and (Di, TV) rc. Then M = Im-D2Y and MDi = Z)2iV from which it follows that Di = D2(N + YDi) = D2U.A\so,XDi -Ni = -N2NmdX = A^2i'which imply that A^i = A^2(A^ + YDi) = N2U. From Di = D2U and the fact that degDi = degD2, it now follows that U is unimodular • The reader should verify that in this case, the relations between the states (3.14) and (3.17) are given by Z2 = Uzi andzi = t/-iz2, i.e., A^ = u,Y = 0,X = U-\
mdE = 0. A similar result is true for IPMFDs given by Di(q)zi(t) = NiiqMt),
1,2.
y(t) = Zi{t\
(3.21)
Specifically, the IPMFDs given in (3.21) are equivalent if and only if there exists a unimodular matrix U such that (3.22)
[Di,Ni] = U[D2,N2l
The proof of this result is completely analogous to the proof of Theorem 3.3 and is omitted. State-space descriptions We now consider the state-space representations given by Xiit) = AiXiit) + Biu{t),
y(t) = CiXi + Di{q)u{t),
i = 1, 2, (3.23)
and we assume that these are related by a similarity transformation, i.e., there is a g ^ ^ n x « |g| ^ 0 , such that Q 0
0 \qln - Ay In\ [ -Ci
By Di(q)_
qln -A2 -C2
B2 \ \Q D2{q)\ [0
0
/„
(3.24)
Clearly, this is a relation of the form (3.3) with M = N = Q 2ind X = Y ^ 0. Therefore, {Ai, Bi, Ci, Di(q)}, {A2, B2, C2, D2(q)} are equivalent in the Fuhrmann or Rosenbrock sense. The converse is also true. In particular, if two state-space representations are equivalent in the sense of Definition 3.1, then they are related
559 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
560 Linear Systems
via a similarity transformation. The proof of this result is not presented here. We ^^^ the reader to verify that in this case the relations between the states (3.14), (3.17) are given by X2 = Qxi and xi = Q~^X2, i.e., N = Q,Y = 0,X = Q-\ and E = 0. Note that the state-space representations considered in (3.23) are more general than the state representations of previous chapters, since the term D{q)u{t) may contain derivatives of the input. In fact it can be shown that every PMD {P{q), Q(q), R(q), W(q)} in (3.1) is equivalent to a state-space description of the form {A, B, C, D(q)} (see, e.g.. Section 2.2 of [30]). Note that a state-space description of the form {A, B, C, D) that is equivalent to a given PMD as in (3.1) exists only when that PMD has a proper transfer function H{s). Now recall the relation between a state-space realization {Ac, Be, Cc, Dc] in controller form of the transfer function matrix H{s) and the corresponding rPMFD {DR, I, NR}. Specifically, (see Subsection 3.4D of Chapter 3 and the Structure Theorem), we have H{s)
with
NR(S)
= CcS(s) +
=
DCDR(SI
NR(S)DR\S) (SI
- Ac)S(s) =
BCDR(SI
(3.25)
where the n X m matrix S(s) = blockdiag [(1, s,..., s^^~^)^\ with dt, i = 1 , . . . , m, the controllability indices of (Ac, Be) or the dimensions of the subblocks in the controller form (Ac, Be). The representations DR(q)z(t) = u(t), y(t) = NR(q)z(t) and xdt) = Acx(t) + Bcu{t), y(t) = Ccx(t) + Dcu(t) are equivalent since Be Dc
0 DRiq) Ip\ y-NRiq)
Im 0
ql - Ac -Cc
Bc\ \Siq) Dc\ [ 0
0
(3.26)
^m
with (Be, ql - Ac) Ic and (DR(q), S(q)) re (show that this is true). Verify that in this case the relation between the states given by Eq. (3.14) is ^^(0 = S(q)z(t), and determine the relation between the states given by Eq. (3.17) for this case. We note that (3.25) can be written as NR(S)
= CS(s) +
DDR(S\
(SI
- A)S(s) =
BDR(SI
(3.27)
where A = Q~^AcQ,B = Q~^Bc,C = CcQ,D = Dc with \Q\ ¥^ 0 and S(s) = Q~^S(s). Equation (3.27) relates NR(S) and DR(S) to a controllable realization of H(s) that is not necessarily in controller form. The matrix 2 is a similarity transformation matrix. Completely analogous results exist for a state-space realization {Ao, Bo, Co, Do} in observer form and the corresponding IPMFD {DL, NL, Ip}Finally, note that the above results involving the Structure Theorem are also valid after the obvious modifications, when the state-space description is of the more general form {A, B, C, D(q)}.
B. Controllability, Observability, Stability, and Realizations Consider now the PMD P(q)z(t) = Q(q)u(t),
y(t) = R(q)z(t) + W(q)u(t\
where P((^) E Riql^^'K Q(q) G R[q]^'''^, R(q) E R[qy\
(3.28)
2indW(q) G /?[g]^x^.The
important system properties of controllability, observability, and asymptotic stability can be developed directly in a manner similar to that in earlier chapters, using state-space descriptions. However, instead of reintroducing these properties, we shall concentrate on establishing criteria for (3.28) to be controllable, observable, and asymptotically stable. Assume that the PMD given in (3.28) is equivalent, in the sense of Definition 3.1 and (3.3), to some state-space representation x(t) = Ax(t) + Bu(t\
y(t) = Cx{t) + Du{t\
(3.29)
where A G /^'^x^ B G R'''''^, C G 7^^x^ and D G T^^^^. That is, for the discussion that follows, we restrict the PMD in (3.28), at least initially, to the class of PMD that is equivalent to state-space descriptions {A, B, C, D] that have been studied extensively throughout this book. For the more general case of (3.28) being equivalent to state-space descriptions of the form {A, B, C, D(q)}, refer to the remarks following Theorem 3.6. Controllability Recall from Section 3.2 of Chapter 3 that {A, B, C, D} is controllable if and only if 1. rank [sil - A,B] = n for any st complex number. 2. The Smith form of [si - A, B] is [/, 0]. Now if {A, B, C, D} in (3.29) is controllable, then in the equivalent description [P, Q, R, W} given in (3.28), the Smith form of [P, Q] will be [/, 0] and rank [P(si\ Qisi)] = I for any complex number Sj, in view of Theorem 3.2(ii). Notice that these conditions are precisely the conditions for P, Q to be Ic polynomial matrices (see Theorem 2.5 in Section 7.2). These observations give rise to the following concept and result. DEFINITION 3.2. The representation {P, Q, R, W} given in (3.28) is said to be controllable if its equivalent state-space representation {A, B, C, D} given in (3.29) is state controllable. THEOREM 3.4. The following statements are equivalent: (i) (ii) (iii) (iv)
{P, e, R, W} is controllable. The Smith form of [P, Q] is [/, 0]. rank [P(si), Q(si)] = I for any complex number Sf. P, garelc.
Proof, Parts (ii) and (iii) follow from the fact that (A, B) is controllable if and only if the Smith form of [ql - A, B] is [/, 0], or equivalendy, rank [siI - A, B] = n for any complex number (see Section 3.2 of Chapter 3), and from Theorem 3.2(ii) in this section. In particular the Smith form of [P, Q] in (ii) is [/, 0] if and only if the Smith form of [ql — A, B] is [/, 0], in view of Theorem 3.2(ii). This in turn is true if and only if (A, B) is controllable, which is equivalent to (i) by definition. So (ii) is true if and only if (i) is tme. Similarly, one can show that (iii) is true if and only if (i) is true. Part (iv) follows easily from (ii) or (iii), in view of Theorem 2.4 in Section 7.2. • Remarks 1. From the above results it follows that (A, B) is controllable if and only if ^/ - A and B are Ic polynomial matrices.
561 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
562 Linear Systems
2. If P, Q are not Ic, then the Smith form of [P, Q] is [A, 0], where A = diag [6/] ¥= I. It can easily be shown that if GL is a geld of [P, Q] = GL[P, Q ] , then the Smith form of GL is A. The roots of the invariant polynomials €( of [P, Q] or of GL are the uncontrollable eigenvalues of the system. 3. Similar tests as the eigenvalue/eigenvector tests of Subsection 3.4B of Chapter 3 can also be developed for the representation {P, Q, R, W}. 4. The rPMFD, {DR, Im, NR}, is clearly controllable since DR and / are Ic. This fact justifies the alternative notation {Dc, /, Nc} for an rPMFD when the controllability of the representation is emphasized; here Dc = DR and A^^ = NR are viewed as matrices of an internal system representation. Note that the term rPMFD stresses the relation to the transfer function H(s) = NR(S)D^^(S), where NR and DR are intrepreted as the numerator and denominator of a transfer function, respectively. Observability Observability can be introduced in a manner completely analogous to controllability. This leads to the following concept and result. DEFINITION 3.3. The representation {P, Q, P, W} given in (3.28) is said to be observable if its equivalent state-space representation {A, B, C, D} given in (3.29) is state observable. • THEOREM3.5. The following statements are equivalent: (i) {P, 2, P, W} is observable. (ii) the Smith form of (iii) rank
\P(Si)
[R(Si)\ (iv) P, R are re.
= / for any complex number st.
Proof, The proofs of these results are completely analogous to the proof of Theorem 3.4 and are therefore omitted. • Remarks . From the above results it follows that (A, C) is observable if and only if ql - A and C are re polynomial matrices. 2. If P, R are not re, then the Smith form of
IS
, where A = diag [e^] T^ /.
It can easily be shown that if GR is a gcrd of
GR, then the Smith
form of GL is A. The roots of the invariant polynomials 6/ of
or of
GR
are
the unobservable eigenvalues of the system. 3. Similar tests as the eigenvalue/eigenvector tests of Subsection 3.4B of Chapter 3 can also be developed for the representation {P, Q, R, W}. 4. The IPMFD, {DL, NL, Ip}, is clearly observable since DL and Ip are re. This fact justifies the alternative notation {Do, No, Ip} for an IPMFD when the observability of the representation is emphasized; here Do = DR and A^^ = NR are viewed as matrices of an internal system representation. Note that the term IPMFD stresses the relation to the transfer function H(s) = DI^(S)NL(S), where NL and DL are interpreted as the numerator and denominator of a transfer function, respectively.
Stability
563
DEFINITION 3.4. The representation {P,Q,R,W} given in (3.28) is said to be asymptotically stable if for its equivalent state-space representation {A,B,C,D} given in (3.29) the equihbrium x = 0 of the free system i = Ax is asymptotically stable. • THEOREM 3.6. The representation {P,Q,R,W} is asymptotically stable if and only if Re ?ii <0,i = 1,..., n, where Xi,i= 1,..., n, are the roots of det P{s)\ the Xi are the eigenvalues or poles of the system. Proof, In view of Theorem 3.2(ii), det {si — A) = a det P{s) for some nonzero a ^ R. At this point it is of interest to briefly discuss controllability, observability, and stability for the more general case, when the PMD in (3.28) is equivalent to a statespace representation of the fonnx{t) =Ax{t) +Bu{t)^y{t) = Cx{t) +D{q)u{t) instead of (3.29). First, note that the concepts of state controllability and observability in state-space descriptions of the form {A,5,C,D(^)} are completely analogous to the {A,5,C,D} case. Furthermore, the criteria for controllability and observability are exactly the same for the above cases, and they depend only on (A,5) and {A^C), respectively. In yiow of this, it can be shovv^n that Definitions 3.2 and 3.3 and Theorems 3.4 and 3.5 for controllability and observability, respectively, are valid for the more general PMD. For similar reasons. Definition 3.4 and Theorem 3.6 on asymptotic stability are valid for the general PMD (3.28). Hovv^ever, care should be exercised vv^hen discussing input-output stability of a system {A,5,C,D(^)} or of their equivalent descriptions of the form (3.28), but this topic vv^ill not be addressed here. For an extended discussion of controllability, observability, and stability of systems described by a general PMD (3.28), refer to Chapter 3 of [9] and Chapter 6 of [33]. It is of interest to note that equivalent representations not only have exactly the same eigenvalues, but also the same invariant zeros, decoupling zeros, and transmission zeros (see Section 3.5 of Chapter 3 for the definitions of zeros and also the discussion belovv^). These assertions follovv^ directly from Theorem 3.2. Poles and zeros In the follovv^ing, wt vv^ill find it useful to first recall the definitions of poles and zeros introduced in Section 3.5 of Chapter 3 and to shovv^ how these apply to PMDs. To this end, consider the representation in (3.28) vv^ith transfer function matrix H{s)=R{s)p-\s)Q{s)+W{s). Let the Smith-McMillan form of H{s) be given by SMH{S)
v^here A(^) = diag
=
0
0
(3.30)
— — , • • • , — — ,r = rank H{s),ei{s) divides e/+i(^),/
\l/r{s) / = 1,..., r — 1 [see (5.3) in Chapter 3]. Then 1,..., r — 1, and 1/A/+1\\l/l{s) (s) divides i/A/(^), 1, and \i/i^i{s) divides \l/i{s)J • the poles ofH(s) are the roots of the characteristic or pole polynomial PH{S) ofH(s) defined as PH{s)
= XI/i{s)---XI/r{s).
(3.31)
Note that PH{S) is the monic least common denominator of all nonzero minors of H{s).
CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
564
It is straightforward to show that
Linear Systems
(3.32)
{poles oiH(s)} C {roots of detP(s)}.
The roots of det P are the eigenvalues or the poles of the system {P, Q, R, W} and are equal to the eigenvalues of A in any equivalent state-space representation {A, B, C, D]. Relation (3.32) becomes an equality when the system is controllable and observable, since in this case the poles of//are exactly those eigenvalues of the system that are both controllable and observable. Comidtr iht system matrix or RosenbrockMatrix of \htvepr:QS&rA&iion{P, Q, R, W}, Sis) =
-Ris)
Q(s) Wis)
(3.33)
and an equivalent state-space representation {A, B, C, D} given in (3.29). In view of (3.5), we have
M X
0 I p^
h-i
0
0
h-n
0
P
Q
0
0
-R
w
0
: o]
0 si
-A -C
:
B \N
-Y
[0
^m
(3.34)
: D\
where M,N E. R[s^^^^ are unimodular and k > max {deg \P\, n) (see (3.5)). The zeros of {P, Q, R, W} can now be defined in a manner completely analogous to the way they were defined for state-space representations. Following the develop \sl - A B~ ment in Section 3.5, let fc = n a n d r = rank , where n r < min(p + -C D 0 n,m-\- n). Consider all those rth-order nonzero minors of Se{s) = \h-i that [ 0 S{s) are formed by taking the first n rows and n columns of Se{s). The zero polynomial of the system {P, Q, R, W}, Zs(s), is defined as the monic gcd of all those minors. The roots of Zs(s) are the zeros of the system {P, Q, R, W}. The reader is encouraged to show that this definition is consistent with the definitions given in Section 3.5. The invariant zeros of the system are the roots of the invariant zero polynomial that is the product of all the invariant factors of S{s). The input-decoupling/outputdecoupling and the input-output decoupling zeros of {P, Q, R, W} can be defined in a manner completely analogous to the state-space case. For example, the roots of the product of all invariant factors of [P(s), Q(s)] are the input-decoupling zeros of the system; they are also the uncontrollable eigenvalues of the system. Note that the input-decoupling zeros are the roots of detGiis), where GL{S) is a geld of all the columns of {P{s), Q{s)\ = GL(S)[P(S), Q(S)]. Similar results hold for the outputdecoupling zeros. The zeros ofH(s), also called the transmission zeros of the system, are defined as the roots of the zero polynomial ofH(s), ZH(S) =
ei{s)...6r(s\
(3.35)
where the 6/ are the numerator polynomials in the Smith-McMillan form of H(s) given in (3.30). When {P, Q, R, W} is controllable and observable the zeros of the system, the invariant zeros, and the transmission zeros coincide.
Consider the representation DRZR = u,y = NRZR with DR G R[S]^^^ and NR E R[sy^^ and notice that in this case the Rosenbrock Matrix (3.33) can be reduced via elementary column operations to the form '
DR
_-NR
I] r / 0\ [-DR
0] \0
I
0
I\ [l
0
r^R
/]ro /" oj [/ 0
7 0
0 -NR
In view of the fact that the invariant factors of S do not change under elementary matrix operations, it is clear that the nonunity invariant factors of S are the nonunity invariant factors of NR. Therefore, the invariant zero polynomial of the system equals the product of all invariant factors of NR and its roots are the invariant zeros of the system. Note that when rankNR = p ^ m, the invariant zeros of the system are the roots of detGi, where GL is the geld of all the columns of NR, i.e., NR = GLNR. When A^^^, DR are re, the system is controllable and observable. In this case it can be shown that the zeros ofH {= NRD^^), also called the transmission zeros of the system, are equal to the invariant zeros (and to the system zeros of {DR, I, NR}) and can be determined from NR. In fact the zero polynomial of the system, Zs(s), equals ZH(S), the zero polynomial of H, which equals ei(s)... er(s), the product of the invariant factor of NR, i.e.. Zs(s) = ZH(S) =
ei(s)...er(s).
(3.36)
The pole polynomial of H(s) is PH(S) =
kdetDR(s),
(3.37)
where k G R. Realization theory and algorithms A representation {P, Q, R, W} is a realization of a rational function H{s) if the transfer function of {P, Q, R, W) is H{s), i.e., if R{s)p-\s)Q{s)
+ Vi^(^) = H(s).
(3.38)
The realization theory for PMDs is analogous to the theory for state-space descriptions, developed in Chapter 5. Results concerning existence and minimality can be developed in the obvious way, using the results on equivalence of representations developed in Subsection 7.3A and the above results on controllability and observability. The following theorems provide results that correspond to Theorems 3.9, 3.10, and 3.11 of Section 5.3 of Chapter 5. The reader is encouraged to give full proofs of these results. THEOREM 3.7. A realization {P, Q, R, W} of H(s) of order n = deg |P| is minimal (irreducible, of least order) if and only if it is both controllable and observable. Proof The proof is left as an exercise.
•
THEOREM 3.8. Let{P, Q, R, W}md{P_, Q, R, W} be realizations of//(^). If {P, Q, R, W} is a minimal realization, then {P, g, R, W} is also a minimal realization if and only if the two realizations are equivalent. Proof. The proof is left as an exercise.
•
THEOREM 3.9. Let {P, Q, R, W} be a minimal realization of H(s). Then the characteristic polynomial of H(s), PH{S), is equal to detP(s) within a multiplication by a nonzero real number, i.Q., PH(S) = kdet P(s). Therefore, the McMillan degree of//(^) equals the order of any minimal realization.
565 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
566 Linear Svstems
Proof, To prove this result, refer to the proof of Theorem 3.11 in Chapter 5 and use the results on equivalence given in Subsection 7.3A. • Realization algorithms It is rather straightforward to derive a realization of H in PMD form. In fact realizations in right (left) PMFD form were derived in Chapter 5 as a step toward determining a state-space realization in controller (observer) form (see Subsections 5.4B and 5.4D). However, these realizations, of the form {DR, Im, NR} and {DL, Ni, Ip], are typically not of minimal order, i.e., they are not controllable and observable. This implies that the controllable realization {DR, Im, NR}, for example, is not observable, i.e., DR, NR are not re. Similarly, the observable realization {DL, NL, Ip} is not controllable, i.e., DL, NL are not Ic. To obtain a minimal realization, a gcrd must be extracted from DR, NR, and similarly, a geld must be extracted from DL, NL. This leads to the following realization algorithm that results in a minimal realization {D, Im, N} ofH. A minimal realization of the form {D, N, Ip} is obtained in an analogous (dual) manner. Consider H(s) = [nij(s)/dij(s)], i = I,..., p, j = 1 , . . . , m, and let lj(s) be the (monic) least common denominator of all entries in the jth column of H(s). Note that lj(s) is the (monic) least degree polynomial divisible by all dij(s), i = 1 , . . . , p. Then His) can be written as H{s) =
(3.39)
NR(S)DR\S),
where NR(S) = [nij(s)] and DR(S) = diag(li(s),..., lm(s)). Note that nij/lj(s) = nij{s)ldij{s) for / = I,..., p, and all j = I,..., m. Now DR(q)ZR(t) = u(t),
y(t) = NR(q)zR(t)
(3.40)
is a controllable realization of H(s). If DR, NR are re, it is observable as well, and therefore, minimal. If DR and NR are not re, let GR be a gcrd and let D = DRG^^ andA^ = NRG^\ Thm D(q)z(t) - u(t),
y(t) = N(q)z(t)
(3.41)
is a controllable and observable, and therefore, minimal realization of H(s) since D, I and D, N are Ic and re polynomial matrix pairs, respectively. Note that ND~ ^ = (NRG^'XDRG^')-'
= (NRG^'XGRD-^')
= NRD-^'
= H.
There is a dual algorithm which extracts an IPMFD resulting in H{s) =
(3.42)
DI\S)NL{S),
which corresponds to an observable realization of H{s), given by DL(q)ZL(t) = NL(qMt),
y(t) = ZL(t\
(3.43)
The details of this procedure are completely analogous to the procedure that led to (3.39). If DL, NL are not Ic, let GL be a geld and let D - G^^DL and N = G^^NL. Then a controllable and observable, and therefore, minimal, realization of H(s) is given by D{q)z(t) = N(qMt),
y(t) = z(t\
(3.44)
In Subsection 5.4B, a controllable state-space realization that is equivalent to (3.40) was obtained first using the controllable version of the Structure Theorem. Next, the unobservable part of this state-space realization was separated and a con-
trollable, observable, and minimal realization was extracted (see Example 4.3 in Subsection 5.4B). This minimal state-space realization is of course equivalent to the realization (3.41), which is in P M F D form. It is possible to extract an equivalent minimal state-space realization directly from (3.41). However, to use the convenient Structure Theorem, typically D has to be reduced first to a column proper form, i.e., (3.41) must be reduced to D{q)z{t)
= u{t),
y{t) =
(3.45)
N(q)z(t\
where D = DU is column proper, N = NU, and f/ is a unimodular matrix. Analogous results are valid for P M F D and realizations (3.42) through (3.44). The following example illustrates the realization algorithms. s^ + 1 ^ + 1
EXAMPLE 3.2. We wish to derive a minimal realization for H(s)
c3
' s^ s Note that this is the same H(s) as in Example 4.3 in Section 5.4B, where minimal state space realizations were derived. The reader is encouraged to compare those results with the realizations derived below. We shall begin with a controllable realization. In view of (3.39)
s^Ji
and H
[s^ + 1, ^ + 1]
= NRD-J,
s^
0" ^
n
c3
. Therefore,
J^RZR = u and y = NRZR constitute a controllable realization. This realization is not ro 01 DR(S) 1 < m = 2, i.e., DR and NR are observable since rank = rank 0 0 NR(S), _i 1. not re. Another way of determining that DR and NR are not re would have been to observe that deg det D(s) = 5 = order of the realization {DR, /, NR}. NOW the McMillan degree of i/, which is easily derived in the present case, is 3. Therefore, the order of any minimal realization for this example is 3. Since {DR, I, NR} is of order 5 and is controllable, it cannot be observable, i.e., DR and NR cannot be re. We shall now extract a gcrd from DR and NR, using the procedure described in Subsection 7.2D. We have 0
DR NR,
"1 5 + 1 —> 5^ 0 0 5.0 5^ Li 5 + 1. "1 5 + r 1 s+V 0 s^ 0 s^ .0 0 J 0 s^ J
c3
S^ + 1
5+1
c3
S+ 1 c3
Therefore, GR = NRG^^,
1 0
using DR =
5+ 1 is a gcrd. We now determine D = DRGJ^^ and A^ = 52 52 - ( 5 + 1) 1 0 5+1 _0
0
5^
1 1 , 5 + 1] = [5^ + 1, - ( 5 + 1)] 0 {DR,I,NR}
0
S
0
=
^
5 S+\
s' -(^+1)
0
s'
=
DGR
and
NR
= [s^ +
NGR, and we verify that they are re. Then 1 0
0 ,[q^ + 1
h-(q+\)]
is a minimal realization of H(s). To determine a minimal state-space realization from Dz = u and y = Nz using equivalence of representations, if so desired, we notice that D is already in column proper form. Using the Structure Theorem, and the notation used in Examples 4.3
567 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
568
and 4.5 in Subsection 5.4B, we obtain D(s)
Linear Systems 1 0
- 1 s^ 1 0
0 s
0
1 0 1 s 0 . Therefore, B^ = D,^ = 0 0 1
u 0 -—1J 11
0
0
1 and Am = 1
1 Also, Dc = lim,^oc//W = [1,0] and A^(^) = [s^ + 1, -{s + 0 "1 0^ -(s+l) 1)] = CcS(s) + DcD(s) = [1,0,0] s 0 + [1,0] Therefore, a minis .0 IJ mal state-space realization of H(s) is given by -Du'Di
=
0 0
Iu + 0
-(s + 1) = DhA(s) + DiS(s) = s
0 0
{Ac, Be, Cc, Dc} =
"0 0 .0
1 0" "0 0" 0 1 , 1 1 ,[1,0,0],[1,0]^ 0 Oj .0 1_
Note that this realization is precisely the minimal realization derived in Example 4.3 in Subsection 5.4B. This will not occur in general, as can be seen if for example a slightly different GR is used. The realization {Ac, Be, Cc, Dc} determined here is in controller form, since D^ was an upper triangular matrix with ones on the diagonal. In general, Dh will just be nonsingular, since D(s) will be column proper, and therefore, the resulting controllable realization will not be in controller form, since Be will not be of the appropriate form. This point was also discussed in Subsection 5.4B and illustrated in Example 4.5 in that subsection. Alternatively, given //, we shall first derive an observable realization. In view of (3.42), H = DI^NL
= (s^y^[s(s^ + 1),^ + 1].
Here Di^(q) and Niiq) are Ic and therefore D(q)z(t) = N(q)u(t) and y(t) = z(t) with D(q) = Di{q) and A^(^) = Niiq) is controllable and observable, and is therefore, a minimal realization. Note that the order of this realization is deg det DL{S) = 3, which equals the McMillan degree ofH{s). In Example 4.3 in Subsection 5.4B a minimal state-space realization in observer form was derived from D{q) and A^(^). •
C. Interconnected S y s t e m s Interconnected systems, connected in parallel, in series, and in feedback configurations, are studied in this subsection. Polynomial matrix and transfer function matrix descriptions of interconnected systems are derived, and controllability, observability, and stability questions are addressed. It is shown that particular interconnections may introduce uncontrollable, unobservable, or unstable modes into a system. Feedback configurations, as well as series interconnections, are of particular importance in the control of systems. Feedback control systems are studied at length in the next section. Systems connected in parallel Consider systems S\ and ^2 connected in parallel as shown in Fig. 7.1 and let
and
Pi{q)ziit)
= Qi(q)um
yi(t)
= Ri(q)zi(t)
P2(q)z2(t)
= Q2(q)u2(t\
y2(t) = R2(q)z2it)
+ Wi(q)ui(t)
(3.46)
+ ^2(^)^2(0
(3.47)
be representations (PMDs) for S] and ^2, respectively. Since u(t) = u\{t) = U2(t)
569 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems FIGURE 7.1. Systems connected in parallel
and y(t) = yi(t) + y2(t), the overall system description is given by Pi(q) 0
0 Piiq)
Zi(q) Z2(q)\
Qi(q) Qiiq).
y(t) =
u(t),
[R,(q),R2(q)]
zi(t) ziit)
+ [WM) +
(3.48) ^2mu{t\
If the systems ^i and S2 are described by the state-space representations i/ = Atxi + BiUu yi = CiXi + DiUu i = 1, 2, then the overall system state-space description is given by i i ] _ \AI X2\ [0
0 1 r^i A2J [^2.
Bi Bi
(3.49)
+ [Di -h D2]u.
y = [Ci, C2]
If Hi(s), H2(s) are the transfer function matrices of 5i and ^2, respectively, then the overall transfer function can be found from y(s) = yi(s) + y2(s) = Hi(s)ui(s) -h H2(s)u2(s) = [Hi(s) + H2(s)]u(s)toht (3.50)
H(s) = Hi(s) + H2(s\
Now if both Hi(s) and H2(s) are proper, then H(s) is also proper. The stability, controllability, and observability of the overall system S represented by Pz = Qu, y = Rz+ Wu in (3.48) will now be examined briefly. In view of |P| = |Pi| IP2I it is clear that the overall system is internally stable if and only if each of the systems 5*1 and ^2 is internally stable. Also, if both Si and S2 are BIBO stable, i.e., all poles in Hi and in H2 have strictly negative real parts, then the overall system is also BIBO stable. The converse is also true, i.e., if the overall system S is BIBO stable, then so are Si and S2. This can be seen from Fig. 7.1 since if, say, ^i were not BIBO stable, then neither would S. Next, this result is also shown using alternative arguments. All uncontrollable or unobservable eigenvalues of Si and ^2 will be uncontrollable or unobservable eigenvalues of the overall system S. To see this, consider Pi 0
0 P2
Qi Q2
and let Gi, G2 be gelds of [Pi, Qi\ and [P2, Q2], respectively. Then
(3.51) Gi 0
0 G2 IS an
570 Linear Systems
Id of the matrix (rows) in (3.51), i.e., the uncontrollable eigenvalues of 5i and ^2 will be uncontrollable eigenvalues of the overall system. Similarly, it can be shown that the unobservable eigenvalues of Si and 52 will be unobservable eigenvalues of the overall system. Note that the overall system S may have additional uncontrollable and unobservable eigenvalues, as is now shown. It is easier to show this when 5i and 52 are controllable and observable. Assume then that = uiit),
yiit) = Ni(q)zi{t)
(3.52)
D2iq)Z2it) = U2(t),
yiit) = N2(q)Z2(t)
(3.53)
DMziit) and
are descriptions for 5i and 52, with//i (5) = Ni(s)D^^^(s)andH2(s) = N2(s)D2^(s). Let Diiq),D2iq) mq]"""" and Ni(q),N2{q) e R[qV'"". The overall system is then described by Di(q) 0 Note that
0 zi(0 D2iq) Z2(t) 0 0
£»2
/ 10 [0
y{t) = miqiN2m zi(t) Z2(t),
uitx 0 / -D2
0] 0 l\
Di 0
-D2 0
(3.54)
Now if for some com-
plex value A rank[Di(X), -D2(A)] < m, i.e., there are fewer than m linearly independent columns in [ A (A),-D2(A)], then ran^t P^^^^^ -^2(A) /„ < 2m, which impHes that A is an uncontrollable eigenvalue of the overall system. If G is a geld of [Di, -D2], then A is a root of detG and it is an eigenvalue that is common in both systems Si and S2. Note that the uncontrollable eigenvalues in G cancel mH = Hi+H2= NiD^^ + A^2^2 ^ = (^1 + A^2^2 ^^O^F^ = (A^i + Completely analogous A^2^2 ^ i ) ^ r ' where D^^Di = (GD2)~^GDi = D^'D results can be shown concerning the unobservable eigenvalues by considering the representations Di(q)zi(t) and
=
Ni(q)ui(tl
yi(t) = Zi(t)
(3.55)
D2(q)z2(t) = N2(q)u2(tl
y2(t) - Z2(t)
(3.56)
with Hi(s) == D^^(s)Ni(s) and H2(s) = D2^(s)N2(s), The unobservable eigenvalues cancel in Di(s)D2^(s) (show this). In view of the above it can be seen that if all the poles of H have negative real parts, then all the poles of both Hi and H2 also have negative real parts, since possible additional poles of Hi and H2 are in the same locations as some of the poles of H. Therefore, if S is BIBO stable so are both 5i and ^^2. Systems connected in series (or in cascade or in tandem) Consider systems ^i and S2 connected in series, as shown in Fig. 7.2 and let Pi(q)zi(t)
= Qi(q)ui(t),
yi(t) = Ri(q)zi(t) + Wi(q)ui(t)
(3.57)
be a polynomial matrix representation for Si and P2(q)z2(t) - 22(^)^2(0.
yiit) = R2(q)z2(t) + ^2(^)^2(0
(3.58)
be a representation for ^2. Here U2(t) = yi(t). To derive the overall system description, consider ^2^2 = Q2U2 = G2J1 = 62(^1^1 + Wiui) and Pizi = Qiui and also y2 = R2Z2 + W2U2 = R2Z2 + W2yi = R2Z2 + ^2(^1^1 + WiwOand j i =
571 ^1
t/2
Vi
Si
CHAPTER 7 :
1 ^^
^2
FIGURE 7.2. Systems connected in series
RiZ\ + WiMl. Then Pi -QiRi
0 Pi
(3.59)
yi = WtPx, /?2] If an external input r2 is introduced as in Fig. 7.3, that is, M2 = Ji + '2» then P2^2 = 22^2 = Qi{y\ + ^2) = QiiRiZi + Wi^O + 22^2 and y2 = R2Z2 + ^2^2 = ^^2^2 + W2(yi + r2) = /^2Z2 + ^2(^1^1 + WiUi) + ^2^2. In this case a more complete description of the two systems in series with inputs wi, r2 and outputs y2, yi is given by Pi -22^1
0 \zi P2 [Z2_
\ «2i [22^1
\yi = Ri W2R1 [y2.
0 ] lui 2 2 J U2.
(3.60)
0] \zi] + Wi \W2WI
Ril U2J
0 W2
If the systems Si, S2 are described by the state-space representations xt = Atxt 4CiUi, yi = CiXi + DiUi, i = 1, 2, then it can be shown similarly that the overall system state-space description is given by Xi
=
hi = J2.
B2C1 Ci D2C1
0] Ixi + Bi A2J [X2_ B2D1
0] «il Bil 7 2 J
0] Ixi + ' D,
0" \ui
C2J [X2_
P2D1
(3.61)
JO2J U2.
If Hi {s), H2(s) are the transfer function matrices of 5i and ^2, then the overall transfer function };2(^) = H(s)ui(s)is His) = H2(s)Hi(s).
(3.62)
It can be shown that if both Hi and H2 are proper, then H is also proper. Note that poles of Hi and H2 may cancel in the product H2H1, and any cancellation implies
FIGURE 7.3. Systems connected in series
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
572
Linear Systems
that there are uncontrollable/unobservable eigenvalues in the overall system internal description. This is discussed further next. The stability, controllability, and observability of the overall system S described by Pz = Qu, y = Rz+ Wu, and given in (3.60) will now be examined. In view of I P| = I Pi 11P21 it is clear that the overall system is internally stable if and only if each of the subsystems Si and ^2 is stable. If both 5*1 and S2 are BIBO stable, i.e., if all poles in Hi and H2 have strictly negative real parts, then the overall system is also BIBO stable. The converse is not necessarily true since cancellations may take place in the product H2H1, i.e., the unstable poles of Hi and H2 may cancel in H2H1 = H, thus leading to a BIBO stable system where Si and/or ^2 might not be BIBO stable. Concerning controllability and observabiHty of (3.60), we make several observations: 1. All eigenvalues of Pi are uncontrollable from r2 and all eigenvalues of P2 are unobservable from yi. This is of course as expected, since Fig. 7.3 reveals that the input r2 does not affect system 5*1 at all, and observation of the output yi will not reveal any information about system 5*2. 2. All uncontrollable eigenvalues of ^i and 5*2 are uncontrollable from ui. This is true because any geld of [Pi, Qi] and any geld of [P2, Q2] are elds of ~ Pi 0 Qi ~ (show this). -Q2R1 P2 Q2W1 Finally, all unobservable eigenvalues of Si and ^2 are unobservable from y2. This is true because any gcrd of Pi and any gcrd of P2 are crds of ^1
Pl -Q2RI W2RI
0 P2 (show this). R2
It should be noted that other uncontrollable and/or unobservable eigenvalues may exist. These correspond to poles of Hi and Hj cancelling in the product H2H1 and can be found from representation (3.60) using, say, the eigenvalue tests for controllability and observability. It is of interest to determine these additional uncontrollable and unobservable eigenvalues directly from the PMFD of Si and 52. To simplify the analysis, we first assume that both Si and 52 are controllable and observable. Let Di(q)zi(t)
= uiit),
yiit) = Ni(q)zi(t)
(3.63)
be a description for 5i, with Hi{s) = Ni{s)D]^^{s), and let D2(q)Z2it) = U2{t),
y2(t) = N2iq)Z2(t)
(3.64)
be a description for 52, with H2(s) = N2is)D2\s). Let Di(q) £ Riqr""", Ni(q) G Rlqy'"" and Diiq) E R[q]P''P, N2(q) G RlqY^'P and recall that yi and U2 have the same dimensions (yi = U2). In this case description (3.60) becomes Di -A^i
0 D2
I 0
0 /
0
0 N2
(3.65)
Di 0 / [-Ni D2 0 Using elementary column operations corresponding to postmultiplication by a uni3stmultip] 0 0 /" , a geld of which is given modular matrix, this matrix is reduced to -Ni £>2 0 The uncontrollable eigenvalues from ui are the roots of a geld of
/ 0 , where GL is a geld of [-A/^i, Di]- Thus, all the uncontrollable eigen0 GL\ values of the overall system are the roots of the determinant of a geld of [-iVi, Di]. Note that these are poles of//2 cancelling in the product H2H\ = N2D2^N\D^^, in D2^N\, or equivalently, in H2N\. The unobservable eigenvalues may be obtained similarly, using the representations by
= Ni{q)ui(t),
yxit) = zi(t)
(3.66)
D2(q)Z2(t) = N2{q)u2{t),
yiit) = ziit)
(3.67)
DM)zi{t) and
with Hi(s) = Dl\s)Ni(s) in this case Di -N2
0 \zi Dil U2.
and H2(s) Ni N2
D2 ^(s)N2(s)- Description (3.60) becomes
Ol \ui A^2j V2_ '
y\ J2.
/ Ol \z\ 0 l\ L^2
(3.68)
Dx 0 The unobservable eigenvalues from y2 are the roots of a gcrd of -N2 62 . Us0 / ing elementary row operations, corresponding to premultiplication by a unimodular Dx 0" GR 0 matrix, this matrix is reduced to -N2 0 , a gcrd of which is given by 0 / 0 / where GR is a gcrd of Dx . Thus, the unobservable eigenvalues of the overall sys-N2 Dx . Note that these are poles of tem are the roots of the determinant of a gcrd of -N2 Hx cancelling in the product //2//1 = D2^N2D\^Nx-> in N2D^^, or equivalently, in NiHx. Systems connected in feedback configuration Consider systems ^i and ^2 connected in a feedback configuration as shown in Fig. 7.4A, or equivalently, as in Fig. 7.4B. Let Pi(q)zi(t) and
= Qx(q)ux(t),
yx(t) = Rx(q)zx(t) + Wx(q)ux(t)
(3.69)
P2(q)z2(t) = 62(^)^2(0,
J2(0 = R2(q)z2(t) + ^2(^)^2(0
(3.70)
be polynomial matrix representations of 5*1 and ^2, respectively. Since ui(t) - y2(t) + rx(t),
(3.71)
U2(t) = yi(t) + r2(tl
(3.72)
where rx and r2 are external inputs, the dimensions of the vector inputs and outputs, ux and j2 and also U2 and j i , must be the same, i.e., ux, y2 ^ R^ and U2, yx ^ ^^• To derive the overall system description, we consider yx = i^iZi + WxUx = ^iZi + ^1(3^2 + n) = Rxzx + Wx(R2Z2 + W2U2) + Wxrx = RxZx + WXR2Z2 + WxW2(yx + ^2) + Wxrx, from which we obtain (/ - WxW2)yx = R\zi^
WxR2Z2 + Win + WxW2r2.
(3.73)
573 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
574 Linear Systems
''^
^1 • > !
- ^
yi
Si
.,
S2
J
(JH
'^
L
FIGURE 7.4A Feedback configuration
^1
+
'L..
-^0^
S2
+
-^H(>-
yi
Sl
FIGURE 7.4B Feedback configuration
Similarly, we have yi = R2Z2 + W2U2 = R2Z2 + ^ 2 ( ^ 1 + ^2) = R2Z2 + ^ 2 ( ^ 1 ^ 1 + Wiui) + W2r2 = R2Z2 + W2RiZ\ + ^2^1(3^2 + n ) + ^ 2 ^ 2 . from which we obtain (/ - W2Wi)y2
= R2Z2 + W2R1Z1 + W2Wiri
+ W2r2.
(3.74)
LEMMA 3.10. det(I - W1W2) = det(I - W2W1). Proof, det(I - W1W2) = = det \1 = det 1 = det { = det Now assume that det(I
I .0 I -Wx
0 I-WxW2_ ) 01 r l\
I
-W2I
0
/ J
/
W2I \i
/J
r/
I --W2WI 0
W2]r / I J [-Wi
0" = det(I /.
- W1W2) = det(I
-W2
0' /. -W2W1)
- W2W1) ¥^ 0. If Mi = (/
-
^ 1 ^ 2 ) ' ^ and M2 = {I - ^ 2 ^ 1 ) - ^ then (3.73) and (3.74) imply that yi = Z\ + [MiWi,MiWiW2] a n d j 2 = [M21^2^!. ^ 2 ^ 2 ] + Z2 r\ . Now Piz\ = Q\ux = Qi(y2 + n ) and P2Z2 = Q2U2 = [M2W2Wi,M2W2\ r2 62(^1 + ^2), where y\ and j 2 are as above. Then the closed loop is described by [MiRi,MiWiR2]
P i - Q1M2W2RX -Q2M1R1 MiRi M2W2R1
P2 -
Q1M2 Q2M1W1
-Q1M2R2 Q2M1W1R2
M1W1R2 M2R2
+
MxWi M2W2W1
Q1M2W2 Q2M1
M1W1W2 M2W2
(3.75) where the identity / + (/ - W i ^ 2 ) ' ^ V^i^2 = (/ - ^ 1 ^ 2 ) " ^ was used
At this point it is of interest to point out that when Wi = 0,W2 closed-loop description (3.75) is simplified to the P M D ^1
-62^1
-QxRi] Ui Pi \ 1.^2.
0 n Qil ['•2. ^
Qi
[o
Ri
y\ J2.
[o
= 0, then the
0 Ui [zi
R2\
CHAPTER 7 :
(3.76)
Note that given system descriptions (3.69) and (3.70), it is always possible to obtain equivalent representations with Wi ^ 0 and W2 = 0. Assuming this is the case, (3.76) is a general description for the closed-loop system, and the condition for the closed-loop system to be well defined is that ^1 -Q2R1
del
-Q1R2] 7^0. Pi
(3.77)
If this condition is not satisfied, then the closed-loop system cannot be described by the polynomial matrix representations discussed in this chapter. If the systems ^i and S2 are described by the state-space representations ki = AiXi + BiUi, yt = CfXi + DiUi, i = 1, 2, then in view of (3.75), the closed-loop system state-space description is Xi
^
U2J Jl
[3^2]
Ai + B1M2D2C1 B2M1C1 MiCi M2D2C1
B1M2 B2M1D1
B1M2C2 A2 + B2M1D1C2
M1D1C2 M2C2
+
M1D1D2 M2D2
M2D2D1
B1M2D2 B2M1 (3.78)
where Mi = (I - DiD2)~^ and M2 = (I - D2DiyK It is assumed that det(I D1D2) = det(I D2Di)y^0. It is not difficult to see that in the case of state-space representations, the conditions for the closed-loop system state-space representation to be well defined is det(I - D1D2) ^ 0. When Di = 0 and D2 = 0, then (3.78) simplifies to Xi
y\ J2.
=
Ai B2C1
Bi B1C2 + 0 A2 _ [X2_
0] B2\
Ci 0 ] Ui\ 0 C2J U2J
(3.79)
EXAMPLE 3.3. Let 5*1 and ^2 be described by zi = u\,y\ = qz\ Sindqzi ~ ^2,y2 = Z2, i.e., {Pi, Gi, Ru Wi} - {1, 1, q, 0} and {P2, Q2, Ri, W2} = {q, 1, 1, 0}. These systems have transfer functions H\{s) = s and H2{s) = l/s, i.e., an ideal inductance Si is connected via feedback to an ideal capacitance S2. Then (3.76) assumes the form ' 1 -q
-1] \zi q \ U2. >i"
yi.
\ 0 q P
575
01 r^i i j U'2_ 01 \z\ Ij U2.
1 -1 0, this does not constitute a well-defined closed-loop descrip-q q tion. It is of interest to note that Si cannot be described by a state-space description of the form {Ai, Bi, Ci, Di] since its transfer function Hi{s) = sh not proper. • Since det
EXAMPLE 3.4. Consider systems ^i and S2 connected in a feedback configuration s+\ s+2 with//i(^) = —-^-^ and//2(5') = and consider the realizations s+2 s+ 1
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
576 Linear Systems
[Pi, Qh R\, Wx] = {q + Zq+1,1, (3.76) becomes q+2 -(q + 2)
0} and {P2, Qi, Ri, W2} = {q+\,q+
(q + 1)1 \zi q+ I \ U2. >i"
J2.
2,1, 0}. Then
'q+ 1 0 1 r^i _ 0 q + 2\ L^2. "1 01 \zi~ 0 ij U2.
q+2 -(^+1)1 = 0, the present example is not a well-defined poly-(q + 2) q+l \ nomial matrix system description. It is of interest to note that here state-space realizations Since det
ofH\ and H2 exist. Let H\ =
+ I, H2 = 7 + 1, and consider the state-space s+2 s -\- I realizations {Ai, Bi, Ci, Di} = {-2, - 1 , 1, 1} and {A2, B2, C2, D2} = {-1, 1, 1, 1}. Here I - D1D2 = 0, and therefore, a state-space reahzation of the closed-loop system does not exist, as expected. •
EXAMPLE 3.5. Consider systems 5*1 and S2 in a feedback configuration with Hi(s) = and//2(^) = 1 and consider the realizations {Pi, Qi, Ri, Wi} = {q+l, q, 1, 0} and [Pi, Qi, Ri, W2} = {1, 1, 1, 0}. Then (3.76) becomes
q+ 1 -1
-q\ ki 1 J U2. >i'
J2.
q 01 In [0 Ij [ri "1 01 \zi [0 ij [z2
+ 1 -q = 1 7^ 0, this is a well-defined polynomial matrix description -1 1 for the closed-loop system. Note that the transfer function matrix of the closed-loop sys1 01 p + 1 -s^ \s 0]\s s tem is H(s) , which is not proper whereas Since det
0
ijL - 1
1J
[0
ij ~ [^ ^ + 1
Hi and H2 were both proper. If the realizations to be considered are {Pi, g i . Pi, W\} = {q + 1 , - 1 , 1, 1} and {P2, Qi, Ri, W2] = {1, 0, 0, 1}, then 1 - WiTy2 = 1 - 1 • 1 = 0 and no representation of the form (3.75) can be derived. This stresses the point that the conditions for the PMD of the closed-loop system to be well defined are given by (3.77), since the condition det (I - W\W2) y^ 0 may lead to erroneous results. Note that det (I - W1W2) = 0 does not necessarily imply that a well-defined polynomial matrix representation for the closed-loop system does not exist. Nowif state-space realizations of ^i(>s) =
- + lmidH2(s) = 1 are considered, s +I namely, {Ai, Pi, Ci, D^} = {-1, 1, - 1 , 1} and {A2, P2, C2, D2} = {0, 0, 0, 1}, then 1 D1D2 = I - I ' I = 0, i.e., a state-space description of the closed-loop does not exist. This is to be expected since the closed-loop transfer function is nonproper and as such cannot be represented by a state-space realization {A, B, C, D}. • Next, let Hi(s) and H2(s) be the transfer function matrices of S\ and ^2, i.e., yi{s) = Hi(s)iii(s)sindy2(s) = 7^2W^2W-In view of wi = })2 + n a n d w 2 = j^i + r2, we have yi = H\U\ = H\{y2 + n ) = H1H2U2 + / ^ i n = H\H2y\ + HiH2f2 + H\ f\, or (I-HiH2)yi
= H,H2r2 +
H,n.
(3.80)
Also, y2 = H2U2 = HziSi + h) = H2H1U1 + H2r2 = H2Hiy2 + / ? 2 ^ i n + ^2^2, or {I-H2Hx)y2
=
H2H,h+H2h.
(3.81)
Assume that det{I - H1H2) = det{I - H2H1) ¥= 0. Note that the proof of the fact that the determinants are equal is completely analogous to the proof of Lemma 3.10. Then
(/ - / / l i / 2 ) - ' / / l H2H\ ( / - H2H\)
tin
Hn
H21
^22
H1H2 (/ H\H2) (/ -H2Hi)-'H2
(3.82)
The significance of the assumption det(I - H1H2) # 0 can be seen as follows. Let Dizi = Niu\y y\ = z\ and D2Z2 = ^2, 3^2 = ^2Z2 be representations of the systems ^i and 52. As will be shown, the closed-loop system description in this case is given by (D1D2 - NiN2)Z2 = A^i^i + 5 i r 2 and yi = D2Z2 ~ ^2 and y2 = N2Z2- Now note that / - H1H2 = I - D\^NxN2D:^^ = p\\DxD2 NxN2)D:^\ which implies that det (/ - H1H2) 7^ 0 if and only if det {D1D2 - N1N2) ¥^ 0, i.e., if det(l - H\H2) = 0, then the closed-loop system cannot be described by the polynomial matrix representations discussed in this chapter. Thus, the assumption that det (I - H1H2) 7^ 0 is essential for the closed-loop system to be well defined. EXAMPLE 3.6. Consider Hi{s) = 2 and H2(s) = 1/s as in Example 3.3. Clearly, 1 H1H2 = 0, which implies that the closed-loop system is not well defined. This agrees with the result of Example 3.3, where it was shown that a PMD of the closed-loop system does not exist. • 5+1 5+ 2 ^,/i^2 = 7 as in Example 3.4. Here 1 5+ 2 5+1 H1H2 = 0 and the closed-loop system is not well defined, which agrees with the results in Example 3.4. • EXAMPLE 3.7. Consider Hi(s) =
EXAMPLE 3.8. Consider Hi(s) =
7 , ^ 2 ^ = 1 as in Example 3.5. Here 1 5+1 H1H2 = —-rr ^ ^' ^^^ therefore, the closed-loop system is well defined. Relation 5+ 1 (3.82) in this case assumes the form 'yi'
=
5
5 1
5
5+lJ
a nonproper transfer function that is the transfer function matrix H(s) derived in Example 3.5. • Now consider systems Si and 5*2 with proper transfer function matrices Hi(s) SindH2(s),3ndlQtXi = AiXi + BiUuyi = CtXi^-DiUui = 1, 2, be their state-space representations. LEMMA 3.11. If det(I - D1D2) 7^ 0, then det(I - Hi(s)H2(s)) 7^ 0. Furthermore, in this case / - Hi(s)H2(s) is biproper [i.e., both / - H\{s)H2{s) and (/ - Hi(s)H2(s))~^ are proper rational matrices]. Proof, It is not difficult to see that I -Hi (5)7/2(5) = I-D1D2 + (strictly proper terms), which implies that det (I - D1D2) is the coefficient of the highest degree 5 term in
577 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
578 Linear Systems
the numerator of det{I — Hi(s)H2(s)). Then if det(I — D1D2) 7^ 0, we have det(I Hi(s)H2(s)) ^ 0. Furthermore, nm(/ - Hi(s)H2(s)) = I- D1D2,
(3.83)
which impHes that when det{I - D1D2) ^ 0, then / - Hi(s)H2(s) is a biproper rational matrix, i.e., I - H1H2 and (/ - H\H2y^ are proper. • In Lemma 3.11, we have used the fact that a proper rational matrix H{s) has a proper inverse, e.g., H~^{s) proper, if and only if detD 7^ 0, where D = lim^^oo H(s). This result is based on Lemma 3.12 and Corollary 3.13. LEMMA 3.12. Consider H(s) G R(S)P^P, which can be expressed uniquely as H(s) = H(s) + Q(s), using polynomial division, where H(s) G R(sy^P is strictly proper and Q(s) G R[sy^P. Then H~^(s) is proper if and only if Q~^(s) exists and is proper. Proof, The proof is left to the reader as an exercise.
•
COROLLARY 3.13. If H(s) G R(s)P^P is proper, then H~\s) is proper if and only if detD # 0, where lim,_oo//(5) = D G RP^'P. Proof, D = Q(s) of the above lemma.
•
Note that det(I - Hi(s)H2(s)) 7^ 0 does not necessarily imply that det(I D1D2) 7^ 0. To see this, recall that in Examples 3.5 and 3.8, Hi = ^ ^ ^, H2 = 1, s+ 1 1 , ¥= 0. However, in any state-space realization and 1 - H1H2 = 1 Di = l,D2 — 1, and 1 - D1D2 = 0. In other words, when Hi(s) and H2(s) are proper, the condition det(I - D1D2) = det(I - D2D1) ¥- 0 implies that det{l Hi(s)H2(s)) = det(I - H2(s)Hi(s)) 7^ 0, or that a closed-loop representation in state-space form exists. When det(I - Hi(s)H2(s)) # 0, a polynomial matrix representation for the closed-loop system exists, but it is not necessarily in state-space form, i.e., det(I — D1D2) is not necessarily nonzero. LEMMA 3.14. Let//i(5), i/2W be proper with Di = \mis-^o,H\{s\D2 = lim^-^ooi/aW and assume that (/ - Hi{s)H2{s)y^ exists. Then Hn{s) = (/ - Hi{s)H2{s)y'^Hi{s) is proper if and only if det (I - D1D2) 7^ 0. Proof If det (I - D1D2) 7^ 0, then in view of Lemma 3.11, / - Hi(s)H2(s) is biproper and Hu(s) is then proper. If Hii(s) is proper, then Hu(s)H2(s) and / + Hii(s)H2(s) are proptv.But(I-Hi(s)H2(s)y^ = I+ (I - Hi(s)H2(s)y^Hi(s)H2(s) = I + Hn(s)H2(sX and therefore, (/ - Hi(s)H2(s)y^ is proper or / - Hi(s)H2(s) is biproper. This implies, in view of Corollary 3.13, that det(I - D1D2) 7^ 0. • Similarly, it can be shown that all the rational matrices in (3.82), namely. yi
Hn
tin \h
h.
H21
H22\ U2.
Ul - HiH2y^Hi (I-H2Hi)H2Hi
(/ - HiH2y^HiH2] In (I-H2Hiy^H2 J U2
are proper if and only if det (I — D1D2) 7^ 0. This result can also be seen from the following system theoretic argument: The rational matrices in (3.82) exist and are proper if and only if there exists a state-space representation of the closed-loop system. This will happen if and only if det (/ - D1D2) 7^ 0, in view of the assumptions in (3.78). In the following it is assumed that det {I - Hi(s)H2(s)) # 0, that is, the closedloop system is well defined.
Controllability and observability Controllability and observability of closed-loop systems are studied next. In view of the representation (3.76), the following is evident: 1. If GIL is a geld of Pi, Qi and G2L is a geld of P2, Q2, then
\GiL 0
0 G2L\
IS a
0 Thus, all uncontrollable eigenvalues of 0 22 ^1 in Gil are uncontrollable eigenvalues of the closed-loop system from ri, eld of
Pi -Q2R1
or r2, or
-Q1R2 Pi
Qi
. Also, all uncontrollable eigenvalues of 52 in G2L are uncontrol-
lable eigenvalues of the closed-loop system from ri or r2, or
. There may be
additional uncontrollable eigenvalues in the closed-loop system, and these are studied below. 0 2. If GiR is a gcrd of Pi, Ri and G2R is a gcrd of P2, R2, then \GiR IS a 0 G2R\ 0 Ri -Q1R2 Pi That is, all unobservable eigenvalues of crd of 0 -Q2R1 P2 R2 Si in GiR are unobservable eigenvalues of the closed-loop system from j i , or J2, or
yi . Also, all unobservable eigenvalues of ^2 in G2R are unobservable
LJ2
eigenvalues of the closed-loop system from j i , or y2, or
. There may be
additional unobservable eigenvalues in the closed-loop system, and these are studied below. Controllability and observabiHty can of course be studied directly from the PMDs and (3.76) as was done above. However, further insight is gained if the additional uncontrollable and unobservable eigenvalues are determined from the PMFDs of ^i and ^2. For simplicity, assume that both ^i and ^2 are controllable and observable and consider the following representations: For system 5i: (la)
Di(q)zi(t)
= ui(t),
or (lb)
Di(q)zi(t)
- Ni(q)ui(t),
yiit) = Ni(q)zi{t)
(3.84)
yi(t) = zi(t),
(3.85)
where (Di(q), Ni(q)) are re and 0i{q), Ni(q)) are Ic. For system 5*2: (2a)
D2(q)z2(t) = U2(t),
or (2b)
j2(0 = N2(q)z2(t)
(3.86)
J2(0 =
(3.87)
D2(q)z2(t) = N2(q)u2(t\
Ut\
where {D2{q), N2{q)) are re and (D2(q), N2{q)) are Ic. In view of the connections uiit)
= y2(t) + ri(t),
(3.88)
U2(t) = yi(t) + r2(t)
the closed-loop feedback system of Fig. 7.4 can now be characterized as follows [see also (3.76)]: 1. Using descriptions (la) and (2a), Eqs. (3.84) and (3.86), we have £>i Ni
-Ni] \z\ £>2j U2.
/
Ol In
0 /J Vl. '
y\
Ni
yi.
0
Ol n (3.89) Nil U2
579 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
580
2. Using descriptions (lb) and (2b), we have
Linear Systems
A^il \z\ D2J [h.
ol ['•1
'Ni
0
A^2j
l.''2. '
I
yi y2.
0
ol \zi i\ U2.
(3.90)
3. Using descriptions (lb) and (2a), we have -I
NxN2\ \zi D2 \ U2.
Ol \n
"A^i
l\ L'"2_ '
0
7
yi 72.
0] r^i
0 7V2J [^2_
(3.91)
Also,_D2Z2 = M2 = >'i + r2 = Dj WiMi + r2 = £)j Wi()'2 + '"1) + r2 = D~[^N\{N2Z2 + n ) + ^2 and yi = M2 — r2 = D2Z2 "" ^2, from which we obtain D2 Z2 + N2
(D1D2 - NiN2)z2 = [Ni,Di]
n (3.92)
4. Using descriptions (la) and (2b), we have
Dx _-A^2A^i
-11 \z\ JO2J
U2.
=
I 0
0 A^2j
V2.'
[y2.
=
Ol \zi 1 0 l\ [Z2.
(3.93)
Also,^Z)izi = ui = y2 + n = D2^N2U2 -^ n = ^2 ^^2(^1 + ^2) + n = £^2^N2(NiZi + r2) + n and 3^2 = ^1 "" n = Dizi — n , from which we obtain 'Ni' 0 Ol nl = Pi zi + {I)2£>l - N2Ni)Zl = [ D2, A^2] ' / oj /2J .3'2j (3.94) The preceding descriptions of the closed-loop system are of course equivalent. This can be proved directly, using the definition of equivalent representations discussed in Subsection 7.3A. Therefore, these descriptions have the same uncontrollable and unobservable modes. In the following analysis we shall carefully select different representations to study different properties in order to secure additional insight. The reader is encouraged to derive similar results, using different representations. It is stressed that in the following, both S\ and 5*2 are assumed to be controllable and observable. The uncontrollability and unobservability discussed below is due to the feedback interconnection only. As was previously shown, there will in general be additional uncontrollable and unobservable eigenvalues in a closed-loop system if S\ and S2 are uncontrollable and/or unobservable. To study controllability, consider the representation (3.89) in (1). It is clear from the matrices
-N2 and D2
that the eigenvalues that are uncontrollable
from ri will be the roots of the determinant of a geld of[—Ni,D2\, and the eigenvalues that are uncontrollable from r2 will be the roots of a geld of [D\, -A^2]- The closed-loop system is controllable from
Clearly, all possible eigenvalues that ^2
are uncontrollable from r\ are eigenvalues of ^2. These are the poles of//2 =^ ^2^2 ^ that cancel in the product H2N1 or, as can be shown, they are the poles of H2 that
/ H\. Similarly, all possible eigenvalues that are uncontrollable from r2 H2 are eigenvalues ofS\. These are the poles ofH\ = NiD^^ that cancel in the product Hi Ho H1N2 or, as can be shown, they are the poles of Hi that cancel in I To study observability, consider the representation (3.90) in (2). It is clear Di -Ni that the eigenvalues that are unobservfrom the matrices and cancel in
able from ^yi will be the roots of the determinant of a gcrd of
, and the eigenD2 values that are unobservable from y2 will be the roots of the determinant of a gcrd of Di . The closed-loop system is observable from . Clearly, all possible eigen-N2 values that are unobservable from yi are eigenvalues of S2. These are the poles of H2 = D2^N2 that cancel in the product N1H2, or as can be shown, they are the poles of H2 that cancel in Hi [I, H2\- Similarly, all possible eigenvalues that are unobservable from y2 are eigenvalues of ^ i . These are the poles of Hi = D^^Ni that cancel in the product N2H1, or as can be shown, they are the poles of Hi that cancel in H2[Hi,Il / Hi) are the eigenHi values that are uncontrollable from r i ; the poles of H2 that cancel in N1H2 (or in HiU, H2\) are the eigenvalues that are unobservable from yi. The poles of Hi that Summary,
The poles of H2 that cancel in H2N1 (or in
cancel in H1N2 (or in
\Hi'
H2) are the eigenvalues that are uncontrollable from r2\
the poles of Hi that cancel in N2H1 (or in H2[Hi, / ] ) are the eigenvalues that are unobservable from j 2 ' The above results may also be expressed in the following convenient way, which, however, should be used only when the poles of Hi are distinct from the poles of H2, to avoid confusion. In the product H2H1, the poles of H2 that cancel are the eigenvalues that are uncontrollable from r i ; the poles of Hi that cancel are the eigenvalues that are unobservable from ^2- In the product H1H2 the poles of ^ i that cancel are the eigenvalues that are uncontrollable from r2; the poles of H2 that cancel are the eigenvalues that are unobservable from yi. EXAMPLE 3.9. Consider systems Si and S2 connected in the feedback configuration ^+ 1 of Fig. 7.4 and let S\ and ^2 be described by the transfer functions Hi(s) = -, and s- I Hiis) = —,—. For the closed loop to be well defined, we must have 1 - H1H2 = s+b (1 — ai)s^ + (b — ai — ao — l)s — (b -\- ao) s + I ais + ao 1 7^ 0. Note that for {s - \){s + b) s- I s+b 1 - 1 = 0 . Therefore, these ^-^mdl-HiH2 ^+ 1 values are not allowed for the parameters if the closed-loop system is to be represented by a PMD. If state-space descriptions are to be used, let Di = \ims-^o=Hi(s) = 1 and D2 = lims^ooH2(s) = a\, from which we have 1 - D1D2 = 1 - ^1 7^ 0 for the a\ = l,ao = -l,andZ? = l,//2
581 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
582
closed-loop system to be characterized by a state-space description [and also, in view of Lemma 3.14, for the transfer functions in (3.82) to be proper]. Let us assume that
Linear Systems
ai 7^ 1.
The uncontrollable and unobservable eigenvalues can be determined from a PMD such as (3.92). Alternatively, in view of the discussion just preceding this example, we conclude the following, (i) The eigenvalues that are uncontrollable from ri are the poles of H2 that cancel in H2N1 = ais + ao {s + 1), i.e., there is an eigenvalue that is uncons+b trollable from r\ (at - 1 ) only when b = \. If this is the case, - 1 is also an eigenvalue ^+ 1 that is unobservable from y\. (ii) The poles of Hi that cancel in H\ N2 = r (ais + UQ) are the eigenvalues that are uncontrollable from r2, i.e., there is an eigenvalue that is uncontrollable from r2 (at -1-1) only when ao/ai = - 1 . If this is the case, -1-1 is also an eigenvalue that is unobservable from j2•
Stability The closed-loop feedback system is internally stable if and only if all its eigenvalues have strictly negative real parts. The closed-loop eigenvalues can be determined from the closed-loop descriptions derived above. If, for example, (3.76) is con-Q1R2 sidered, then the closed-loop eigenvalues are the roots of det \ Pi [-Q2R1 Pi Since any uncontrollable or unobservable eigenvalues of S\ and ^'2 is also an uncontrollable or unobservable eigenvalue of the closed-loop system, it is clear that for internal stability, both ^i and 5*2 should be stabilizable and detectable, i.e., any uncontrollable or unobservable eigenvalues of 5*1 and ^2 should have negative real parts. This is a necessary condition for stability. We shall nov^ concentrate on the descriptions (3.87) to (3.94) in (1) through (4). First, recall the identities det
A C
D = det(A)det(B B
- CA'^D)
= det(B)det(A
- DB'^C),
(3.95)
v^here in the first expression it v^as assumed that det {A) ^ 0 and in the second expression it was assumed that det(B) 7^ 0. The proof of this result is imme-
I 0] \A D -CA-^ l\ [c B = A-DB-^C 0 C B
diate from the matrix identities [/
[0
- D f i - ' l \A
D
/ J [c B
D B-CA-^D
and
-N2 , (D1D2 D2 -N2 D2 N1N2), and {D2D1 - N2N1) from the closed-loop descriptions in (1), (2), (3), and (4). Then Now consider the polynomial matrices
det
£>! -Ni
-N2 £>2
= det(Di) det{D2 -
NiD^^N2)
= det(Di)det(D2
Dx^N\N2)
-
= det(Di) det01^) det0iD2 - N1N2) = aidet(DiD2
- N1N2),
(3.96)
where at is a nonzero real number. Also
CHAPTER 7 :
-N2 D2
det
583
= det(D2) det(Di - 7V2^2 ^M) = det(D2) det(Di -
D2^N2Ni)
= det(D2) det 02^) det(D2Di - N2N1) = a2 det 02^1
(3.97)
- N2N1I
where 0:2 is a nonzero real number. Similarly, det
-N2
D2
= didet02Di
-
N2Ni\
(3.98)
where di = det(Di) det(D^ ^) is a nonzero real number, and det
-N2
Di
= d2det(DiD2
- N1N2I
(3.99)
where ($2 "= det 02) det(D2^) is a nonzero real number. These computations verify that the equivalent representations given by (1), (2), (3), and (4) have identical eigenvalues. The following theorem presents conditions for the internal stability of the feedback system of Fig. 7.4. These conditions are useful in a variety of circumstances. Assume that the systems Si and 52 are controllable and observable and that they are described by (3.84) to (3.87) with transfer function matrices given by Hi = NiDi^ and
= Di^Ni
(3.100)
H2 = N2D2^ = D2^N2
(3.101)
where the {Nt, Di) are re and the {Nt, Di) are Ic for / = 1,2. Let a 1 (s) and a2(s) be the pole (characteristic) polynomials ofHi(s) and H2(s), respectively. Note that ai(s) = kidet(Di(s)) = kidet0i(s)), i = 1, 2, for some nonzero real numbers ki, ki. Consider the feedback system in Fig. 7.4. THEOREMS.15. The following statements are equivalent: (i) The closed-loop feedback system in Fig. 7.4 is internally stable, (ii) The polynomial (a) det (b) det
Di Ni
-N2
bx
-Ni'
Ni
Di.
O2.
or or
(c) det0iD2 -NiN2),ov (d) det02Di -N2N1) is Hurwitz; i.e., its roots have strictly negative real parts, (iii) The polynomial ai(s)a2(s)det(I - Hi(s)H2(s)) == ai(s)a2(s)det(I - H2(s)Hi(s)) is a Hurwitz polynomial.
(3.102)
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
584
iv) The poles of
Linear Systems
ill
I -Hi
Ml.
-H2 I
-1
^1
72.
(3.103)
{I-H2H1)-' Hi(I-H2Hir
H2(I-HiH2r'] \ri
{I-HIH2V
J U2
are stable, i.e., they have negative real parts. (v) The poles of 'y\ yi.
-1 " 0
'-H2
I
. I
-Hi_
H2]\ri 0 J L^2_
L^l
(3.104)
(I-HiH2r'HiH2]
(I-HiH2r'Hi {I-H2H1)'H2H1
(/ - H2H1)
\ri H2 J ih
are stable. Proof. It was shown previously in (3.96) to (3.98) that all the determinants in (ii) are equal within some nonzero constants, i.e., they have the same roots. These roots are the eigenvalues of the closed-loop feedback system [recall that the representations (3.89) to (3.94) are all equivalent]. The feedback system is internally stable if and only if its eigenvalues have negative real parts. This shows that (ii) is true if and only if (i) is true. T o s h o w ( i i i ) , l e t D ^ = D1D2- NiN2 2indwvitQ det (I - H1H2) = det(D^\DiD2NiN2)D2^) = det(D^^) det(D2^) det(Dk). Since ai(s) = kidet(Di(s)) and a2(s) = k2 det (D2(s)), where ki, k2 are nonzero real numbers, it follows that, ai(s)a2(s) det (/ Hi(s)H2(s)) = ki /c2 det(Dk), which is a polynomial, the roots of which are the closedloop eigenvalues. Note also that det (I - H1H2) = det (I - H2H1). Since the roots of (3.102) are exactly the closed-loop eigenvalues, (iii) is true if and only if (i) is true. To show (iv) we write / -Hi
D2 0
H2 I
0" 61^
-1 r
62
-N2 \Di D2_ [0
Di -Ni
-7^2" Di_
[-Ni
0"
(3.105)
-1
D2_
and we notice that these are coprime polynomial matrix factorizations (show this). The poles of
/ -Hi
-H2 I
are the roots of J^H I
Di -Ni
-N2 D2
that are precisely the closed-
loop eigenvalues. Note that the poles are also equal to the roots of det [ | det
Di
-Ni
-N2
D2
[see (ii) above] since
D2
0 / / 0
-N2
-Ni
Di
D2 -Ni
-N2 Di
bi
-Ni
N2
D2
0 / / 0
So (iv) is true if and only if (i) is true since the poles in (3.103) are exactly the eigenvalues of the closed-loop system. "0 n lui - ~riT , which in view of the proof yi' To show (v), we note that = y2_ 1^2. of (iv) above implies that 'yi
= 0 nl /
I ojl
= 0 /ir
-Hi I
/ oj^ -Hi
-H2 I -H2] I
I 0
ro [HI
0 / H2] 0
/ -Hi
-H2 I
-1
.0
0 ] \b. bi\ [ 0
"0
N2\
_/ OJ
Nx
0 J Vh.
"0 /I To / ] [ 61 J oj 1 / OJ L-A^2
- ^ i l [0 ^2] [/
"0
/]
-N2
D2
/ oj
/]
"0
or
'yi
" A
.h.
,-/!/2
-1
-N2
62
D2
0' Di.
-1
"0
A^2l \r\
.A^i
0 J ih.
585 CHAPTER?:
\r\ • 1 - 1
/] oJ
"0 Ni
Ni] \r\ 0 J ih.
01 Tn-1 -A^if ^\Ni 62J [0 7V2J ih '
(3.106)
which is an Ic factorization. This of course is the expression for the transfer function that could have been derived directly from the internal description in (3.90). Similarly, from (3.89) it can be shown that 0
yi
0
N2j
Di -Ni
-N2
- 1
I
(3.107)
D2\
which is an re factorization. Clearly, the poles of the transfer function matrix in (3.104) that relates
and
are precisely the closed-loop eigenvalues, which implies that
(v) is true if and only if (i) is true.
•
Remarks 1. Parts (iv) and (v) of Theorem 3.15 can also be shown to be true by using the following brief argument. It was shown that the feedback system is controllable from
is observable from or
. Also, it can easily be shown that the system
and observable from
to
. Therefore, the poles of the transfer function from
are precisely the eigenvalues of the closed-loop system, which
must have negative real parts for internal stability. 2. It is important to consider all four entries in the transfer function (3.104) between and
[or in (3.103) between
and
] when considering inter-
nal stability. Note that the eigenvalues that are uncontrollable from r\ or r2 will not appear in the first or the second column of the transfer matrix, respectively. Similarly, the eigenvalues that are unobservable from y\ or j2 will not appear in the first or the second row of the transfer matrix, respectively. Therefore, consideration of the poles of some of the entries may lead only to erroneous results, since possible uncontrollable or unobservable modes may be omitted from consideration, and these may lead to instabilities. 3. Write (/ - HiH2y^
= D2(DiD2 -
( / - -H2Hi)-^
=
NiN2y^Di
(3.108)
DiiD2Di-N
(3.109)
If
Dk = D1D2 - N1N2
(3.110)
and
Dk = D2D1 - N2NU
(3.111)
and
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
586
then (/ - 7/1//2)"' = D2D-'Du
(I -
H2H1)
-- DiDl^D2,
and
Linear Systems D2Dl^Ni N2Dl^Nx
NiDl^N2 DiDl^N2
NiD-^D2 [N2D-^Ni
Note that the first column above can be written as
NiDl^N2 N2D^^Di\
(3.112)
D2 D^^N\, and possible canN2
D2 are re. In view of the N2, representation (3.92), these are precisely the eigenvalues of the closed-loop system that are uncontrollable from r i . Similar results can be derived by considering the second column. These results agree with the comments made in Remark (2) above. Similarly, consider the first row NiDl^[D2,N2\. Since the [^2> A^2] are Ic, possible cancellations may occur only between Ni and D^, which in view of representation (3.94) are precisely the eigenvalues of the closed-loop system that are unobservable from y. Analogous results can be established by considering the second row N2Dl^{N\, D\\. Finally, we note that cellations may only occur between D^ and A^i since the
N2D
DiD ^-1 D2D-'Ni
D^D l^r^2 D2D-'Di
(3.113)
from which similar results concerning uncontrollable and unobservable eigenvalues can easily be derived. 4. The roots of each of the polynomials in (ii) of Theorem 3.15 are equal to the roots of the polynomials in (iii) and are equal to the poles of the transfer functions in (iv) and in (v). They are exactly the eigenvalues of the closed-loop system. 5. The open-loop characteristic polynomial of the feedback system is a\{s)a2{s). The closed-loop characteristic polynomial is a monic polynomial, aci{s), with roots the closed-loop eigenvalues, i.e., it is equal to any of the polynomials in (ii) within a multiplication by a nonzero real number. Then, relation (3.102) implies, in view of (4), that the determinant of the return difference matrix [I—Hi (s)H2(s)] is the ratio of the closed-loop characteristic polynomial over the open-loop characteristic polynomial within a multiplication by a nonzero real number EXAMPLE 3.10. Consider the feedback configuration of Fig. 7.4 with Hi =
s+1
s- r
ais + ao the transfer functions of systems Si and 5*2, respectively. Let ai 7^ I s+b so that the loop is well defined in terms of state-space representations (and all transfer functions are proper). (See Example 3.9.) All polynomials in (ii) of Theorem 3.15 are equal within a multiplication by a nonzero real number, to the closed-loop characteristic polynomial given by aci(s) = 9 b — ai — ao — I b + ao ^, . , . , , ^^ . , . ,„ s + s . This polynomial must be a Hurwitz polynomial for I — ai I — ai internal stability. If ai(^) = s — I and a 2(5) = s + bsiVt the pole polynomials of Hi and H2, then the polynomial in (iii) is given by ai(s)a2(s)(l - Hi(s)H2(s)) = (1 - ai)s^ + (b — ai - ao - l)s - (b -\- ao) = (I — ai)aci(s), which implies that the return difference Ho =
1 -
Hi(s)H2(s)
(1 - ai)
ai(s)a2(a)
-. Note that (1 - H1//2)"
(1 -
H2Hir
(s - m + b) with a(s) = (1 - ai)aci(s), and the transfer function matrix in (iv) of a{s)
Theorem 3.15 is given by
587
(s - l)(s -\- b) (s ~ l)(ais + ao) a(s) a(s) (s + 1)(^ + b) (s - l)(s + b) a{s) a{s) The polynomial a{s) has a factor s + \ when b = I. Notice that a(-l) 2-2b = 0 when b = 1. If this is the case (b = 1), then -s - I (s - l)(ais + ao)a(s) a(s) s- 1 s+ 1 a(s) a(s) where a(s) = (s + l)d(^). Notice that three out of four transfer functions do not contain the pole at - 1 in d(s). Recall that when b = 1, - 1 is an eigenvalue that is uncontrollable from ri, and it cancels in certain transfer functions, as expected (see Example 3.9). Similar results can be derived when ao/ai = -I. This illustrates the necessity for considering all the transfer functions between wi, W2 and ri, r2 when studying internal stability of the feedback system. Similar results can be derived when considering the transfer functions between j i , y2 and ri, r2 in (v). • COROLLARY 3.16. The closed-loop feedback system in Fig. 7.4 is internally stable if and only if (i) Djf i[(/ - H2Hi)-\ H2(I - HiH2)-^] is stable, or (ii) D2^[Hi(I - H2Hi)-\ (I - HiH2y^] is stable, or (I-H2H1)-' D2 ^ is stable, or (iii) [Hi(I - H2H1)-' (iv)
H2(I - H1H2)-' D^ ^ is stable. (I - H,H2)-'
Proof. The above expressions originate from rows and columns of the transfer function matrix between
and
. In view of (3.113), expression (i) is equal to
D\^DiDl^[D2, N2\ = D^^[D2, A^2], the poles of which are precisely the roots of detDk, the closed-loop eigenvalues, since the pair (D2, N2) is Ic. Similar arguments hold for (ii), (iii) and (iv). • It is of interest to point out that expression (i) of the corollary is precisely the transfer function between the partial state z\ and the input and Dizi
, in view of (3.103)
= u\. Expression (ii) corresponds to the transfer function between zi and
. Notice that \ir\ = =
=
(i-H2Hir' (I-HiH2)-'Hi
//2^2 +
\l-H2Hi)-'H2 {I-H,H2)-'
Hxh
from (3.104). Since H2 = ^ 2 ^^2 is an Ic factorization, the transfer function from zi\ to r2 is stable if and only if expression (iii) is stable. A similar argument holds for (iv). The following result can be proved in a completely analogous manner to the proof of the above corollary.
CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
588 Linear Systems
COROLLARY 3.17. The closed-loop feedback system in Fig. 7.4 is internally stable if and only if (i) D f H / - H2Hi)-^D2^ is stable, or (ii) D^^I - HiH2T^b^^ is stable.
•
Note that expression (i) in the corollary is stable if and only if the transfer function between z\ and
is stable. To see this, note that u\ = (I - HjHi)
w h i c h z i = Df^wi = [ Z ) j ~ H / - ^ 2 ^ i ) " ^ 2 D2U1 - N2U2 = [D2,
-N2\
^ n , from
] ^ 2 n . N o w D 2 r i = D2U1 - D2y2 =
,i.e.,£i -
[D-,\l-H2Hi)-'D^'W2,-N2}
Since the pair (62, N2) is Ic, this transfer function is stable if and only if expression (i) is stable. Similarly, U2 = (I - H\H2)~^r2^ from which £2 = [^2 H^ ~ HiH2r'D^']Dir2
= [D^\I
-
HiH2r'D^'][-Ni,D,]
, which is stable if
and only if expression (ii) is stable, since (-A^i, Di) is an Ic pair of polynomial matrices. s+ 1 and 7/2 W = s- 1 a\s + ao , as in Example 3.10. The expressions in Corollary 3.16 are given by s+b {s - \){s + b) (s- l)(ais + ao) 1 1 [s + b, ais + ^o], (i) a(s) a(s) ' a(s) s- 1
EXAMPLES.11. Consider again the systems described by Hi (s) =
(ii)
(iii)
(iv)
(s + l)(s + b) 1 s+b a(s)
(s-
l)(s + b) a(s)
a(s)
[s+hs-
1],
(s - l)(s + b) 1 s- 1 1 a(s) s+ 1 a(s)' (s + l)(s + b) s + b a{s) (s - l)(ais + ap) 1 a{s) ais + ao 1 1 I s + b \ a(s) (s - 1)(^ + b) a(s)
Clearly, the monic pole polynomial of each of the transfer functions in (i) to (iv) is a(s) 9 (b - ai — ao - I) (b + ao) , ^ , = s + s- — = cici(s), the closed-loop characteristic I — ai 1 — a\ \ — a\ polynomial. This is so even when Z? = 1 or aolai = - 1 (see Example 3.9), in which case there are uncontrollable/unobservable eigenvalues. Notice that s + b and a\s + ao are coprime by the definition of H2(s). Similarly, the expressions in Corollary 3.17 are given by 1 (s - l)(s + b) 1 s— I a(s) s+b 1 (s - l)(s + b) 1 (ii) s+b a(s) s- 1 as expected. (i)
1 a(s) 1 a(s)'
PART 2 SYNTHESIS OF CONTROL SYSTEMS
589 CHAPTER 7 :
7.4 FEEDBACK CONTROL SYSTEMS In this section, feedback systems are studied further with an emphasis on stabiHzing feedback controllers. In particular, it is shown how all the stabilizing feedback controllers can be conveniently parameterized. These parameterizations are very important in control since they are fundamental in methodologies such as the optimal /7°°-approach to control design. Our development of the subject at hand builds upon the controllability, observability, and particularly, the internal stability results introduced in Section 7.3, as well as on the Diophantine Equation results of Subsection 7.2E. First, in Subsection 7.4A all stabilizing feedback controllers are parameterized, using PMDs. A number of different parameterizations are introduced and discussed at length. State feedback and state estimation, using PMDs, are introduced in Subsection 7.4B and are then used in Subsection 7.4C, where all stabilizing feedback controllers are parameterized, using proper and stable MFDs. Two degrees of freedom feedback controllers offer additional capabilities in control design and are discussed in Subsection 7.4D. Control problems are also described in this subsection.
A. Stabilizing Feedback Controllers Now consider systems Si and ^2 connected in the feedback configuration shown in Fig. 7.4, or equivalently, in Fig. 7.5. Given S\, it is shown in this section how to parameterize all systems ^2 so that the closed-loop feedback system is internally stable. Thus, if 5*1, called the plant, is a given system to be controlled, then ^2 is viewed as the feedback controller that is to be designed. Here we provide the parameterizations of all stabilizing feedback controllers.
^^O
S^
yi
3^2
Or
FIGURE 7.5A Feedback configuration
'^ /-2
+^-N.
^2 ,
Q ^ >H
Si
yi
FIGURE 7.5B Feedback configuration
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
590 Linear Systems
Several parameters will be used to parameterize all stabilizing systems S2. The results are based on PMDs of the systems S\ and 5*2 and of the closed-loop system, in particular, PMFDs. Parameterizations are introduced, using first the polynomial matrix parameters (1) D^, N^ and D^, Nk and then the stable rational parameter (2) K = NkDl^ = D^^TV^. These parameters are very convenient in characterizing stability, but cumbersome when propemess of the transfer function H2 of 5*2 is to be guaranteed. The rational proper and stable parameters (3) Qi and Q2 are introduced next. These are very convenient when properness of H2 is to be guaranteed, but cumbersome when characterizing stability, except in special cases, e.g., when Hi is stable. The stable parameters (4) Su and 521 that are the comparison sensitivity matrices of the feedback loop will also be used. These can be employed to conveniently parameterize all stabilizing H2 controllers only in special cases, for example, when / / f ^ exists. Parameters (5) X2 and X2, which are closely related to Q2, will also be introduced and discussed. The results based on X2 in the case of SISO systems reduce to well-known classical control results. Finally, rational parameters (6) Li, L2 are introduced to help in providing alternative proofs for the results that involve the previously introduced parameters. These results provide a link with parameterizations of all stabilizing controllers using proper and stable factorizations. Such parameterizations are addressed in Subsection 7.4C. The distinguishing characteristics of our approach of the parameterizations of all stabilizing feedback systems S2 (of all stabilizing controllers) is that they are based on internal representations, which were developed at length earlier in this book. This approach builds upon previous results and provides significant insight that allows the introduction of several parameters that are useful in different circumstances. This approach also makes it possible to easily select the exact locations and number of the desired stable closed-loop eigenvalues. Utilizing the development of this section, a parameterization that uses proper and stable MFDs and involves a proper and stable parameter K' is introduced in Subsection 7.4C. The parameter K' is closely related to the stable rational parameter K used in the approach enumerated above. This type of parameterization is useful in certain control design methods such as optimal //°°control design. In the following, the term "stable system 5" is taken to mean that the eigenvalues of the internal description of system S have strictly negative real parts (in the continuous-time case), i.e., the system S is internally stable. Note that when the transfer functions in (3.103) and (3.104) of the feedback system S are proper, internal stability of S implies BIBO stability of the feedback system, since the poles of the various transfer functions are a subset of the closed-loop eigenvalues (see Section 7.3 and Chapter 6). Feedback systems The feedback system of Fig. 7.5 was studied at length in Subsection 7.3C. Recall that if Pi(q)Zi(t) = Qi(q)uiit) ytit) = Ri(q)Zi(t),
i
=12,
are polynomial matrix descriptions of Si and ^'2, then the closed-loop system is given by (3.76), i.e..
^1
yi = 'J2_
0
591
0]
QIR2] Qx = 0 P2 J [Z2j
Pi -QiRi
U2_
(4.2)
ol \zi\ R2\ [Z2\
where the relations
Mi(0 = nCO + nit),
mit) = yiit) + nit)
(4.3)
have been used. Based on this description, it was shown that for the closed-loop eigenvalues to be stable (to have strictly negative real parts), it is necessary for 5i and ^2 to be stabilizable and detectable. This is so because the uncontrollable and unobservable eigenvalues of both S\ and ^2 will also be uncontrollable and unobservable eigenvalues of the closed-loop system. Now assume that S\ and ^2 are controllable and observable, and let system S\ be described by Niiq)ziit)
(4.4)
= Ni(q)ui(t), yiit) = zi(t),
(4.5)
(la)
Dx{q)Zi{t) = ui(t),yiit)
(lb)
Di(q)zi(t)
=
or
where the pair {D\{q), N\{q)) is re and the pair (D\iq), Ni{q)) is Ic. Let Hi{s) = Ni{s)D^^{s) = D'[^(s)Niis) be the transfer function matrix of S'l. Next, let system 52 be described by (2a)
D2(q)z2{t) = U2(t), yiit) = N2iq)z2(t)
(4.6)
(2b)
D2(q)Z2{t) = N2(q)u2(t), y2{t) =
(4.7)
or Mt),
where the pair {D2{q), N2{q)) is re and the pair {D2{q), N2(q)) is Ic. Let H2is) =
N2(s)D^Hs) D2 ^(s)N2(s) be the transfer function matrix of S2. Recall from Subsection 7.3C [see (3.89) to (3.94)] that the closed-loop system descriptions are in this case as follows: 1. Using descriptions (la) and (2a), we have Di
A^2l \zi
D2\
7 ol \n
U2.
0
\yi
'Ni 0
[y2.
i\ U2 0 ] \zi /^2} L^2.
(4.8)
2. Using descriptions (lb) and (2b), we obtain
-N2
Ni] \zi D2J [Z2. \yi y2.
'Ni
0
ol ['"'
A^2j
[''2.
7 Ol [zi -| 0
/J LZ2
(4.9)
CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
592
3. Using descriptions (lb) and (2a), we have
Linear Systems (D1D2 - NiN2)Z2 = [Ni,Di] D2 N2
(4.10)
Z2 +
4. Using descriptions (la) and (2b), we obtain {D2D1 - N2Nl)Zl = 02, N2]
zi +
0 -/
0 0
(4.11)
These descriptions of the closed-loop system are all equivalent and the polynomials det
-N2
det{DiD2-NiN2),
det and
-N2
-Ni D2
det(D2Di - N2N1)
are all equal within multiplication of nonzero scalars. Their roots are the closed-loop eigenvalues (see Theorem 3.15). Parameterizations of all stabilizing systems S2 The six parameterizations of all stabilizing systems ^2 previously mentioned are now introduced and discussed at length. 1. Parameters Dt, N^ and Dj^, N^. Consider now the closed-loop description given in (4) above and let (4.12)
D2D1 - N2N1 = Dk
where D^ is some desired polynomial matrix so that the roots of detD^ are stable (have negative real parts). If Z)i, A^i, D^ are assumed to be given and D2, -N2 that satisfy (4.12) are to be determined, then this equation is seen to be a polynomial matrix linear Diophantine equation. Such equations were studied in Subsection 7.2E. In view of Theorem 2.15 of Chapter 7, Eq. (4.12) has a solution for any Dj^ (of appropriate dimension) since the pair (Di, A^i) is re. All solutions of the Diophantine Equation (4.12) can be parameterized as follows [see (2.54)]:
or
D2 = DkX, - N^N,
(4.13a)
-N2 = DkYi + NkDi
(4.13b)
[D2,-N2] = [Dk,Nk]
Xx
(4.14)
where X\, Y\ satisfy X\D\ -\- Y\N\ = /, the pair {Ni,D\) is Ic and is such that —N\D\ + D\N\ = 0, and N^ is an arbitrary polynomial matrix of appropriate di\Dx mensions. Note that {D^Xi, D]J^\\ is a particular solution of {D2, -A^2] = D,. ^1
For the system D2Z2 = N2U2, J2 = Z2 to be well defined, we must have detD2 = det (Dj^Xi — NkNi) 7^ 0, i.e., Nk (and D^) must be such that this condition is satisfied. Similar results can be derived using the closed-loop description given in (3). Let (4.15)
D1D2 - N1N2 = Dk,
where D^ is some desired polynomial matrix so that the roots of det D^ are stable. In view of Theorem 2.15, all solutions of the Diophantine Equation (4.15) are given by D2 = XiDk - NiNk
(4.16a)
-N2 = YiDk + DrNk
(4.16b)
N2 D2_
or
Di Ni
-Yi\ \-Nk Xi\ [Dk
(4.17)
where Xi, Yi satisfy DiXi + A^i7i = /, the pair (Ni,Di) is re and is such that —DiN\ + NiD\ = 0, and Nk is an arbitrary polynomial matrix of appropriate diXiDk is a particular solution of [Di, Ni] D2 = Dk. For mensions. Note that YiDk -N2 the system D2Z2 =" ^2, j2 = ^^2^2 to be well defined, we must have detD2 = det{XiDk - NiNk) 7^ 0, i.e., Nk (and Dk) must be such that this condition is satisfied. The preceding two sets of parameterizations for ^'2 are of course related. A convenient way of presenting the above results in a unified way is to use the doubly coprime factorizations of H\. To this end, assume that the descriptions (4.4) and (4.5) for system Si satisfy UU-'
=
Xi
Fi
£»!
A^i A J [A^I
-I'l ^i.
/ 0
0 /
(4.18)
where U is 3. unimodular matrix (i.e., det U is a. nonzero real number) and the X\, Yi,Xi, Yi are appropriate matrices. Factorizations of Hi = NiD^^ = D^^Ni that satisfy (4.18) are called doubly coprime factorizations of Hi. Such factorizations always exist (see, e.g., [21], [1]). In particular, in Theorem 2 of [1] it is shown that if the relations XiDi + FiA^i - /, DiXi + Nifi = I and -NiDi -f- DiNi = 0 are satisfied, then (4.18) is satisfied for Di, A^i, Di, Ni, Xi, Yi and [Xi, Yi] = [Xi, Yi] -h ^[-A^i, Di], where S = XiYi - YiXi, i.e., any Ic and re factorizations of Hi with associated matrices can be adjusted so that they are doubly coprime. If Di, Ni and Di, Ni are doubly coprime factorizations and satisfy (4.18), then the solutions of the Diophantine Equations in (4.14) and (4.17) can be written as [D2.-N2] and
= [Dk,Nk\U
N2 = U- 1 D2
-Nk
(4.19) (4.20)
from which it follows that [Dk,Nk] = and
-Nk Dk
[D2,-N2]U-'
= u\ N2 D2
(4.21) (4.22)
593 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
594
Xhe relation between D^, N^ and D^, N^ is now clear, namely,
Linear Systems
-DuNk
+ NkDk = 0.
(4.23)
THEOREM 4.1. Assume that the system Si is controllable and observable and is described by the PMD (or PMFD) as (i) DiZi = u\, y\ = N\Z\ given in (4.4), or by (ii) b\z\ = N\u\y y\ = zi given in (4.5). Let the pair (Di,Ni) and the pair (Di,Ni) be doubly coprime factorizations of the transfer function matrix//i (5) = MDj"^ = D^^Ni [i.e., (4.18) is satisfied]. Then all the controllable and observable systems S2 with the property that the closed-loop feedback system eigenvalues are stable (i.e., have strictly negative real parts) are described by (i)
D2Z2 = N2U2,
y2 = Z2,
(4.24)
where D2 = DkXi - NkNi and N2 = -{DkYi + A^^^i) with Xi, Fi, A^i, 5 i given in (4.18) [see also (4.19)] and the parameters D^ and Nk are selected arbitrarily under the conditions that D^^ exists and is stable, and the pair {Dky Nk) is Ic and is such that det(DkXi -NkNi)y^O. Equivalently, all stabilizing ^2 can be described by (ii)
D2Z2 = W2,
yi = N2Z2>
(4.25)
where D2 = XiDk - NiNj, and N2 = -{Y\Dk + i^iA^^) with Xu ?i, Nu Di given in (4.18) [see also (4.20)] and the parameters Dk and A^^^ are selected arbitrarily under the conditions that D^^ exists and is stable, and the pair (Dk, Nk) is re and is such that det(XiDk-NiNk)7^0. Furthermore, the closed-loop eigenvalues are precisely the roots of detbk or of detDk. In addition, the transfer function matrix of 5*2 is given by H2 =
-0kX,-NkN,)-\DkY,+Nkb,)
= -{Y,Dk + DiNk)(X,Dk - NiNk)-\ (4.26) Proof, The closed-loop description in case (i) is given by (4.11) in method (4) above and in case (ii) it is given by (4.10) in method (3). As was shown in Subsection 7.2E, (4.19) and (4.20) are parameterizations of all solutions of the Diophantine Equations b2Di - N2N1 = Dk [in (4.10)] and D1D2 - N1N2 = Dk [in (4.11)], respectively. The fact that 5^^ (or D^^) exists and is stable guarantees that the closed loop is well defined and all of its eigenvalues, which are the poles of D^^ (or of D^i), will be stable. The condition det(bkXi - NkNi) ¥^ 0 [or det{XiDk - NiNk) ¥^ 0] guarantees that detb2 7^ 0 [or detD2 ¥^ 0], and therefore, the PMD for ^2 in (4.10) is well defined. Finally, note that the pair 0k,Nk) is Ic if and only if the pair 02, N2) is Ic as can be seen from 02, -N2] = 0k,NkW given in (4.19), where U is unimodular. This then implies that the description 02, N2> 1} for S2 is both controllable and observable. Similarly, the pair (Dk,Nk) is re guarantees that {D2,1, N2} with D2 and N2 given in (4.20) is also a controllable and observable description for 5*2. • 2. Parameter K. In place of the polynomial matrix parameters Dk, N^ or Z)^, Nk, it is possible to use a single parameter, a stable rational matrix K. This is shown next in Theorem 4.2. THEOREM 4.2. Assume that the system Si is controllable and observable and is described by its transfer function matrix Hi = NiD^^ = b^^Ni,
(4.27)
where the pairs (A^i, Z)i), 0i, Ni) are doubly coprime factorizations satisfying (4.18). Then all the controllable and observable systems S2 with the property that the closedloop feedback system eigenvalues are stable (i.e., they have strictly negative real parts) are described by the transfer function matrix
H2 = -(Xi - KNir\Y,
+ KDi)
595
= -(Fi + D,K)(Xi - NiKr\
(4.28)
where the parameter K is an arbitrary rational matrix that is stable and is such that det{X\ - KN\) # 0 or det(Xi — NiK) 7^ 0. Furthermore, the poles of ^ are precisely the closed-loop eigenvalues. Proof, This is in fact a corollary to Theorem 4.1. It is called a theorem here since it was historically one of the first results established in this area. The parameter K is called the Youla parameter (see Section 7.6). In Theorem 4.1, descriptions for H2 were given in (4.26) in terms of the parameters Dk, Nk and Dk, Nk. Now, in view of -DuNk + N^Dk = 0, given in (4.23), we have D-,'Nk = NkDl'
(4.29)
= K,
a stable rational matrix. Since the pair (D^, N^) is Ic and the pair {Nk, Dk) is re, they are coprime factorizations for K. Therefore, H2 in (4.28) can be written as the H2 of (4.26) given in the previous theorem, from which the controllable and observable internal descriptions for S2 in (4.24) and (4.25) can immediately be derived. Conversely, (4.28) can immediately be derived from (4.26), using (4.29). Note that the poles of K are the roots of det Dk or det Dk that are the closed-loop eigenvalues. • s+ 1 EXAMPLE 4.1. Consider//i = . Here A^i = A^i = 5 + 1 and Di = Di = 5 - 1 . s- 1 These are doubly coprime factorizations (a trivial case) since (4.18) is satisfied. We have Di
Xi -Ni
uu-
s+ -{s +1)
-Yi Xi
-s + s-\
s- 1 s+l
-{-s + |) s+ \
In view of (4.26) and (4.28), all stabilizing controllers H2 are then given by {-s+l)dk^{s-\)nk
H2
{s + \)dk -{s+ \)nk
_ _{-s+l) + {s-l)K (5 + i ) - (5 + 1)K '
where K = rik/dk, any stable rational function. EXAMPLE 4.2. Consider//i(5") = '
-(s + 1) 1
[1,0]
= NiD-,1
_
1 [1, s+l] = Di - 1W i , which are coprime polynomial MFDs. Relation (4.18) is given by D,
Xi
uu-^ =-Ni
5i
1
Xi
-s^ + l
s+ 1 s'^ + s + 1
-1
0
-Fi
'2
-(S + 1)
0
1
-(s + 1)
~(s+l)
0
:
1
All stabilizing controllers may then be determined by applying (4.26) or (4.28).
CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
596
We shall now express the transfer functions between yi or J2.
Linear Systems
U2
and /2_
m
terms of the parameters D^, N^ or D/., N^ or K. Recall that [see (3.103) and (3.104)]
\y\ = and
Ui
(I -
(I-HxH2r'HiH2 (I-H2Hi)-'H2
HIHJT'
(I-H2Hi)-'H2Hi
H2(I
=
Hid -
U2
H2Hir'
HiH2r'
(4.30)
(4.31)
It is not difficult to see, in view of (4.26) and (4.28), that ( / - i ? i / / 2 ) " ' =D2D^^Di = (XiDk - NiNk)Dl^Di = (Xi - NiK)Du and (/ - //z^/i)"^ = £>i-D^'^2 = DiiXi - KNi). Also, (/ - HxH2)-^Hi = D2Dl^Nx = Dxbl\bkXx-NkNx) (XiDk - NxNk)D7'Nx (Xi - NiK)Nx = Hiil - H2Hi)-^ = NiDl^D2 = NxDl\DkXi - NkNx) = NxiXx - KNx), and (/ - H2Hx)-^H2 = Z^ID^'JVJ = -DxDl'{DkYx+NkDx) = -DxiYx + KDx) = H2(I - HxH2)-' = N2Dl'Dx = -{YxDu + DxNk)Dl'Dx = - ( f i + DxK)bx. Note that (/ - HiH2)-^HxH2 = Hx{I - H2Hxr^H2 = Nxb, 'N2 = -NxDl\DkYx + A^^^i) = -A^i(Fi + KDx). To express (/ - HxH2)~^HxH2 in terms of D^, N^ we write it as D2D^^NxH2 = -{XxDk-NxNk)Dl'Nx{YxDk + DxNu)0txDu-NxNkr^ = -{Xx-NxK)Nx{Yx + DxK){Xi — NxK)~^, which is not as simple as the expression given in terms of bk,Nk. Similarly, (/-772^i)-iH2Hi = H2{I-HxH2)-^Hx = A^2£>^'A^i = -(?i£>*+ DxNk)Dl'Nx - ( F i + DxK)Nx. To express (/ - //2ffi)~'^2^i in terms of Di,jV<;,wewriteitasDiD^i7^2/^i = -DxDl^{DkYx+Nkbx)NxD:^^ = -DxiYx + KDx)NxDiK It is straightforward to express the transfer functions (4.30) and (4.31) in terms of the parameter ^ , in view of A" = N^D^^ = D^'iV^. In particular, we have J2.
U2
=
(Xi -NiK)Ni - ( f 1 + DxK)Nx
-(Xx - NxK)NxH2 -{Yx + DxK)Dx
(4.32)
=
(I-H2NxD^')(Xx -NxK)Nx
-{Yx + DxK)Dx {Xx- - NxK)Dx
(4.33)
These expressions were derived from the previous expressions given that involve Dj^ and N^. The right factorization of H2 in (4.28) is to be considered in these expressions. Similarly, ^
U2
=
NxiXx -KNx) -DxiYx + KDx)NxD7 DxiXx-KNx) NxiXx-KNx)
-NxiYx +KDx) -DxiYx + KDx)
-DxiYx + KDx) il - b^^NxH2)-^
(4.34)
(4.35)
which were derived from the previous expressions involving D^ and N^. The left factorization of H2 in (4.28) is to be considered in entry (2, 2) of the
to
transfer function. These expressions for the transfer functions in terms of K can of course be combined in order to use the most convenient parameterizations. 3. Parameters Q\ and Q2. Recall from Subsection 7.3C the relations (3.103) and (3.104) given by
and
I -Hi
-1
=
-H2 I
=
(I-HiH2r'Hi (I-H2H,r'H2H,
=
I -til
-H2 I
"0 til
H2]
0 J U2_ (I-H,H2r'H,H2] (I-H2H,r'H2
In J U2_
(4.36)
-1
(I-H2H1)-' Hi{I-H2Hir'
H2{I-HiH2r' (I-H,H2r'
(4.37)
We shall now express the transfer function matrices in terms of some important parameters. To this end, define = (/ -
H,H2)-'H,
(4.38)
Q2 ^ H2Q - HxH2)-' = (/ -
H2H,)-'H2
(4.39)
Hi = Qxil + //2Q1)-' = (/ +
QiH2)-'Qi
(4.40)
Qi = / / i ( / - H2Hi)-'
and note that
H2 = & ( / + HiQ2)-'
={I + Q2H,)-'Q2.
(4.41)
It is not difficult to see that QxH2 = H1Q2 and Q2H1 = H2Q1 from Q1H2 = (I - H\H2r^HxH2 = Hi(I - H2Hi)-^H2 = H1Q2, and similarly for Q2HX. Also note that (/ - //1//2)"' = / + HiQ2 = / + QiH2
(4.42)
(I - H2Hx)-^ = / + H2Q1 = / + 2 2 ^ 1 ,
(4.43)
from which by postmultiplying the first relation by Hi and the second by H2, we obtain Qi = (/ + HiQ2)Hi = Hi(I ^ Q2H1) and 62 = (/ + H2Qi)H2 = H2(I + Q1H2). In view of these relations, it is now straightforward to write yi
[y2_ Ui U2
=
Q\ H2Q1
Q1H2 (I + H2Qi)H2
I + H2Q1 Qi
(/ + H2Qi)H2 1 + Q1H2
(4.44)
(4.45)
Note that (4.44) and (4.45) are useful when H2 is given, in which case they depend only on the parameter g i . In this case Hi is given by (4.40). It is not difficult to see that Qi is the transfer function matrix between yi or U2 and ri. Similarly, \yi
[h. -^ Ui
^ U2
Hi(I + Q2Hi) QiHi I + Q2H1 — Hi(I^Q2Hi)
H1Q2 Q2
(4.46)
Q2 I + H1Q2
(4.47)
597 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
598 Linear Systems
Expressions (4.46) and (4.47) are useful when H\ is given, in which case the transfer functions depend only on the parameter Q2. In this case H2 is given by (4.41). Q2 is the transfer function matrix between J2 or wi and r2. If the given transfer function H2 or Hi is proper, then it is straightforward to guarantee properness of all other transfer functions by appropriately selecting the parameter Qi or Q2. In particular, given that Hi is proper, let a proper Q2 be such that (I -\- HiQ2)~^ exists. Then, in view of Lemma 3.14 in Subsection 7.3C, H2 = Q2{J + HiQ2)~^ given in (4.41) is proper if and only if det{I + (lim5^oo//i)(lim5^oo22)) ^ 0. If this is true, then all the transfer functions in (4.46) and (4.47) will be proper. This implies that if the given Hi is strictly proper then any proper Q2 will guarantee that the above transfer functions exist and are proper. If Hi is not strictly proper, i.e., (lim^^oo//i 7^ 0), then care should be taken in the selection of a proper Q2; if in this case it is possible to select Q2 strictly proper, then H2 will be (strictly) proper for any g2- Similar results are valid when H2 is given and Qi is selected. In Theorem 4.3 we assume that Hi is given and the parameter Q2 is to be chosen to guarantee internal stability. Such a situation arises when Hi represents the given plant and H2 is the stabilizing controller to be designed. It should be noted that (as was shown in Subsection 7.3C), if the given system ^i is not controllable and observable, all its uncontrollable and unobservable eigenvalues will appear as uncontrollable from
and unobservable from
or
eigenvalues in the closed-loop
system. Therefore, for stability it is necessary to assume that Si is stabilizable and detectable. Here, working with Hi, any possible uncontrollable and unobservable parts of Si are ignored. Exactly analogous results exist for the parameter Qi when H2 is given. Consider the feedback system of Fig. 7.5 and assume that Si is controllable and observable with transfer function matrix Hi. THEOREM 4.3. Given//i, let H2 = G2(/ + //lG2)"
(4.48)
where the rational matrix Q2 is such that (/ + HiQ2)~^ exists. Then the eigenvalues of the closed-loop feedback system will be stable (i.e., they will have negative real parts) if and only if Q2 is selected so that (i) the poles of Hi(I + Q2H1), H1Q2, Q2H1, and Q2 are stable, or (ii) the poles of Df^/ + 22^1, G2] are stable, or (iii) the poles of
\
2^ 1b^ ^
are stable.
Furthermore, the eigenvalues of the closed-loop system are precisely the poles in (i) or (ii) or (iii). In addition, if Hi is proper, then H2 and the transfer matrices in (4.36) and (4.37) are proper if and only if Q2 is proper and det (I + (lim^^oo //^(lim^^oo Q2)) ^ 0. Proof, The expressions in (i) are precisely the entries of the transfer matrix between and
given in (4.46), the poles of which are the closed-loop eigenvalues (see
Theorem 3.15 in Subsection 7.3C). The expressions in (ii) and (iii) come from Corollary 3.16, where it was shown that the poles in these expressions are exactly the closedloop eigenvalues. The statement about properness was proved above, just before the theorem. •
599
Remarks 1. It is not difficult to see that this theorem characterizes all H2 that lead to a feedback system with stable eigenvalues. When Hi is proper it also characterizes all proper stabilizing H2. 2. The eigenvalues of the closed-loop system are the poles of the expressions in conditions (i) or (ii) or (iii). These will be poles of Q2, but also stable poles of H\, depending on the choice for Q2. 3. It is easy to select Q2 so that H2 in (4.48) is proper (when H\ is proper). Also, it is not difficult to select Q2 so that the stability conditions are satisfied. However, to conveniently characterize all Q2 that satisfy the stability conditions (i), (ii), or (iii) is not straightforward, unless special conditions apply. This is the case for example when Hi is stable, as in Corollary 4.4. 4. For further discussion of the relation between propemess and stability in a system described by a PMD, refer to Chapter 6 of [331 and also [9]. COROLLARY 4.4. Given //i, assume that its poles are stable. Let H2 be given by (4.48), where the rational matrix Q2 is such that (/ + HiQ2)~^ exists. The eigenvalues of the closed-loop feedback system will be stable (i.e., will have negative real parts) if and only if Q2 is stable. In addition, if Hi is proper, then H2 and all the transfer matrices in (4.36) and (4.37) are proper if and only if Q2 is proper and det{I + (lim,_oo//i)(lim,_oo Q2)) ^ 0. Proof, When the poles of H\ are stable (i.e., have negative real parts), then the conditions (i) or (ii) or (iii) of Theorem 4.3 are true if and only if the poles of Q2 are stable. • Corollary 4.4 greatly simplifies the conditions on Q2. However, it is valid only when Hi has stable poles. The result in this corollary was pointed out by Zames in [38], where it was used to introduce the //°°-framework of optimal control system design. 4. Parameters Sn and 821- At this point it is of interest to introduce the matrices
S12 = (I- HiH2r\
and
52i = (I - / / 2 ^ i ) ' \
(4.49)
which are of importance in control system design. They are the companion sensitivity matrices of the feedback loop. Su is also the transfer function between U2 and r2, while ^21 is the transfer function between ui and ri. Note that Q2 = H2{I HiH2r^ = H2S12 = S21H2 and Qi = Hi(I - i/2//i)"^ = H1S21 = SnHi, Also, H1Q2 = S12 - /, 22^1 = ^21 - / and Q1H2 = S12 -1, H2Q1 = S21 - L Now consider Theorem 4.3 and note that H2 = Q2S 12
(4.50)
^2"/22
and that the conditions (ii) and (iii) can be written as D. ^[821, Q2] and 22 DI ^12
which must be stable in order to have stable closed-loop eigenvalues. This implies that a necessary condition for stable closed-loop eigenvalues is that all unstable poles of Hi must appear as zeros of 821 and of 812. In general it is not possible to parameterize all stabilizing controllers in terms of only 812 or 821, but this can be accomplished under certain assumptions on Hi. In particular, we have the following result.
CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
600
COROLLARY 4.5. Assume that detHx 7^ 0. Let
Linear Systems
H2 = H[\Sn or
- I)S^2 = DdN{\Sn
- I)]Su'
(4-51)
H2 = 52-/(^21 - I)H[' = S2i'[(S2i - I)N{']Du
(4.52)
where ^12 or ^21 are rational matrices such that S^2 ^^^ ^21^ exist. Then the eigenvalues of the closed-loop feedback system will be stable if and only if (i) S12 is such that the poles of
H;\Sn-I) ^12
are stable,
D^'
(ii) ^21 is such that the poles of Z)f^ [521, (52]-/)//f^] = [Dl^S2u D^\S2i are stable. Proof. The proof of this result follows directly from Theorem 4.3.
I)N;^Di] •
It is clear from (i) of Corollary 4.5 that all the unstable zeros of ^ i must cancel with zeros of 5*12 — / , which is the transfer function between yi and r2 [it equals (/ - H\H2)~^H\H2\. Also all the unstable poles of Hi must cancel with zeros ofSn. From (ii) a similar result follows for ^'21 - / , which is the transfer function between j2 and ri [it equals (/ - / / 2 ^ i ) ~ ^ ^ 2 ^ i ] and for 821 [see Exercise 7.23(c) and (d)]. 5. Parameters X2 and X2' Alternative parameters, closely related to Q2, may be used in Theorem 4.3. To this end, let X2 = Di^Q2,
and
X2 = Q2Di\
(4.53)
= [(/ + X2Ni)D^^]-^X2,
(4.54)
COROLLARY 4.6. Given//i, let H2 = DiX2(I + NiX2r^
where the rational matrix X2 is such that (/ + A^iX2)~^ exists. The eigenvalues of the closed-loop feedback system will be stable (i.e., will have negative real parts) if and only if X2 is selected so that the poles of [iI + X2N,)D^\X2]
(4.55)
are stable. Furthermore, these poles are precisely the eigenvalues of the closed-loop system. In addition, if Hi is proper, then H2 and all the transfer matrices in (4.36) and (4.37) are proper if and only if D1X2 is proper and det (I + (lim^-^oo Hi)(\ims^o, D1X2)) 7^ 0. Proof. Assume that the poles in (ii) of Theorem 4.3 are stable. Then, in view of Q2 = Z)iZ2,Dfi[/ + G2^bG2] = Dl^lI + DiX2NiD;\DiX2] = [{I + X2Ni)D^\X2]^\so has stable poles. Conversely, if (4.55) has stable poles, then D]^^[I + Q2H1, Q2] also has (the same) stable poles, where Q2 was selected to be Q2 = D1X2. The remainder of the corollary follows easily in view of Theorem 4.3. Note that X2 can be seen to be the transfer function between z\ and r2, since Q2 is the transfer function between ui = Dizi and r2 and Q2 = D1X2. Also N1X2 = H1Q2 = (I - HiH2Y^H\H2 is the transfer function matrix between y\ and r2. It should be pointed out that contrary to Corollary 4.4, the result in Corollary 4.6 is valid for Hi unstable as well [7]. In a completely analogous manner, it can easily be shown that the following result is true. COROLLARY 4.7. Given//i, let H2 = (1 + X2Ni)-^X2bi
- X2[b-,\I
+ NxX2)r\
(4.56)
where the rational matrix X2 is such that (/ + X2N\Y^ exists. The eigenvalues of the closed-loop feedback system will be stable (i.e., will have negative real parts) if and only if X2 is selected so that the poles of X2 + NiX2).
D^\l
(4.57)
are stable. Furthermore, these poles are precisely the eigenvalues of the closed-loop system. In addition, if Hi is proper, then H2 and all the transfer matrices in (4.36) and (4.37) are proper if and only if ^2^1 is proper and det (/ + (lim^_>oo X2Di)(\ims^oo Hi)) 7^ 0. s-h 1 of Example 4.1. In view of Theorem 4.3, all s- 1 s+ 1 1 where stabilizing controllers H2 are given by i^2 = G2 1 + 1-H
EXAMPLE 4.3. Consider Hi =
&^,a
— stable implies that n Id 1 {s - \)d + (5 - \)n{s + 1 ) d + n{s+ 1) {s - \)n, and d is stable. Also, stable ^- 1 d{s-\) {s- \)d implies that d{\) + n{\)2 = 0 and d is stable. Therefore, for stability Q2 = nid with d stable, and n = {s - l)n, d{l) + n{\)l = 0. In view of Corollary 4.6, all stabilizing controllers H2 are given by H2 = (^ - 1 ) [ 1 + is stable. If Q2 = n/d, then
^2(^+1)]"^^2, where 1 +X2(S+ 1)-
1
,X2 is stable. If X2 = n/d, then d is stable.
and^(l) + /i(l)2 = 0 for internal stability. Note that ^2 = n/d = D1X2 = {s-l)X2 (s - l)nld.
= m
1 -^+ ^ - r [ l , 5 + l] = Dr^TVi, whichis an c2' c2 is"- sIc polynomial MFD. In view of Theorem 4.3, all stabilizing controllers H2 are given 22 5 f ^ are stable. If Q2 by H2 = Q2(l + HiQ2)~\ where the poles of 1 + ^^162 EXAMPLE 4.4. Consider Hi(5)
, then Q2D^^ stable implies that ni = s^fii, ^2 = s^h2 and d is stable. Now (1 + un^f^-\ HiQ2)Di
-
s^ + rii + {s + l)n2 1 1 + m + (^ + l)/22 ^^ ^ = - ^ stable imposes additional
conditions on fii and «2. EXAMPLE 4.5. Consider Hi =
1
-. This system is stable and therefore Corol-
5+ r
lary 4.4 applies. All stabilizing controllers H2 are given by H2 = ^2(1 + ^ 1 6 2 ) ^ = Dllei n ({s + \)d + {s- \)n^~^ -, where Q2 n/d is stable. n(s + 1) d (s+l)d ) {s+l)d + {s-\)n Furthermore, if Q2 is proper with lim^-^oo Q2(^) ^ - 1 , then H2 will be proper. Note that in view of Theorem 4.3, the closed-loop eigenvalues will be the poles of (5 + 1) + (5 - \)n n {s+\)d 'd\
1 -[(s+ l)d + (s(s+l)^d
l)n,(s+
1 s+ 1
l)nl
Relations among parameters. It is now straightforward to derive relations among the parameters Q2 and Qi; X2 and X2; D^, N^, and D^, Nk', K\ and also 512 and 521- In particular, in view of (4.32) to (4.35), we have
601 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
602
(22 = //2(/ - HxH2r^
Linear Systems
= N2D^^Di =
= (/ -
H2Hiy^H2
Dibl^N2
= -(YiDk + D,Nk)Dl^bx = ~Dxbl\bkYi = - ( f 1 + DiK)bi = -£>i(Fi + KDi).
+ NuDi) (4.58)
Also, Qi = Hid - H2H1)-' = (/ = NiDl^b2
H,H2r'Hi
= D2Dl^Ni
= Nibl\bkX, - N„Ni) = (XiDk - NiNk)D-k^Ni = NiiXi -KNi) = (Xi - NiK)Ni.
(4.59)
The parameters Q2 are used when the system Si is assumed to be given. We have K = -D-[\Q2
+ Yrb,)bi'
= -D\\Q2
+ D,Y{)b\\
(4.60)
Next, we note that in view of (4.49) and (4.50), we have 5i2 = / + HxQ2 = / + Nxbl^N2
= / - Nibl\bkYx
+ A^2^i)
= / - A^i(yi + Kbx) and
521 = / + Q2H1
= 1 + N2D^^Ni = I - (YiDk +
DxNk)Dl^bi
= / - (fi + DxK)bx.
(4.61)
(4.62)
From (4.53) it now follows that Q2 = D1X2 = X2D1, from which the parameters X2 and X2 can be expressed as X2 = - ( 7 1 + Kbi)
and
X2 = - ( ? i + DiK).
(4.63)
The expressions of H2 in terms of the parameters are now summarized: H2 = -0kXi
- NkNi)-\bkYi
= -(Xi - KNxr\Yx
+ Nkbi)
= - ( f , £ ) , + DiNk)iXiDk -
+ Kbx) = - ( F i + £>ii^)(^i -
T^XN^Y'
^xKT^
= (/ + Q2Hxr'Q2 = [DfH/ + e2Hi)]-uor'Q2] = Qiii + HxQi)-^ = [e25r>][(/ + i?ie2)5ri]-' = 52-/22 = [D-x'S2xr\D-x'Q2\ = [(/ + X2Nx)Dx^r'X2
= Q2Sx^ = [Q2b-x'}[Sx2b~x'T'
= MbxHl
+ NiX2)r\
(4.64)
6. Parameters Lx, La and Lx, L2. Expressions (4.64) that describe H2 as a ratio of rational matrices can also be derived in an alternative way, using a theorem that we establish next. This result also provides a link between the development of this subsection and the descriptions of all stabilizing controllers using proper and stable factorizations given in Subsection 7.4C. Consider the feedback system of Fig. 7.5 and assume that 5i is controllable and observable. The following constitutes one of the principal results of this section. THEOREM 4.8. Let the transfer function matrix of the system Si be Hi with Hi = NiD^^iHi = D^^Ni) being a coprime polynomial matrix factorization. If H2 is the transfer function of ^2, then the eigenvalues of the closed-loop feedback system are stable if and only if H2 can be written as
//2 = L^^U
(H2 = L1L2 ^),
(4.65)
where the L2, L\ (Li, L2) are stable rational matrices with detL2 # 0 {detLi # 0), which satisfy L2D1 -LiNi
= I
01L2 -NiLi
= I).
(4.66)
DZ
(4.67)
Furthermore, if [L2,Li] = D-,'[D2,N2]
are coprime polynomial matrix factorizations, then the closed-loop description is given by DkZ = 02, N2]
72.
yi J2.
'
In D,z = [Ni, bi]
y\
z+
Pi. D2 N2_
0
01
. - / oj
0 z+ .0
(4.68)
- / I In 0 J1^2.
Proof, The proof of the parts in parentheses will not be shown since it follows the proof given below in a completely analogous manner. Let [Li, L2] be stable with detL2 9^ 0 and satisfying (4.65) and (4.66) and write an Ic polynomial matrix factorization as in (4.67). Then D2D1 - N2N1 = D^, where D^^ exists and is stable. Furthermore, the pair 02, N2) is Ic [since any eld of the pair (D2, A^2) would be an Id of Dk] and ^2 ^ exists. Therefore, the closed-loop system with//2 = ^ 2 ^ ^ ! "^ 62^/^2is well defined. Its internal description is given in (4.68) [see also (4.11)], and the closed-loop eigenvalues are the stable roots of detbk. Now let H2 = 62^ N2 be an Ic polynomial factorization and note that the closed-loop system is given by (4.68), where Dk = D2D1 - N2N. Assume that D^^ is stable. Define [L2, Li] = bl^02, N2] and note the D^ and 02, A^2] are Ic since the pair 02, N2) is Ic Then H2 = ^2 ^^1 where L2 and Li are stable with L2D1 LiNi = / . In view of the above theorem, all stabilizing H2 are given by (4.65), v^here the stable rational matrices L2, Li (Li, L2) span all solutions of the Diophantine Equation given in (4.66). There are different ways of parameterizing all stable solutions L2, Li (L2, Li) of (4.66) with det L2 7^ 0 (det L2 ¥" 0), each one leading to a different parameterization of all stabilizing H^. In fact, in this way one may generate the parameterizations developed above, thus providing alternative proofs for those results. This is shown in the next corollary. COROLLARY 4.9. All Stabilizing H2 are given in the following. (i)
^ 2 ~ ^2
^1
{H2 =
(4.69)
L1L2'),
where the L2, Li (Li, L2) are stable with detL2 7^ 0 (detL2 7^ 0) and L2D1 - L\N\ = / 0\L2
— N\Li = / ) . The closed-loop eigenvalues are the poles of [L2, ^i] ( L VLMJ
(ii)
H2 = -(Xi - KNir'iYi
+ KDO
[H2 = -(Yi + DiK)(Xi -
NiKy'l (4.70)
where i^ is any stable matrix such that J^r(Zi - i^A^i) # 0[det(Xi-NiK) ¥^ 0].TheZi, [7 01 r ^i Yx\ pi - ^ I , where f/ is a unimodular Fi,Xi,Fi satisfy t/t/ -1 _
[-Af,
61J [iV,
Xlj
0
/.
603 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
604
matrix. The closed-loop eigenvalues are the poles of
Linear Systems
Xi - NiK
[Xi - KNi, Yi + KDi]
(4.71)
or equivalently, the poles of K. (iii)
H2 = [D-,\I + Q2H,)r'[D-,'Q2]
= (/ +
QiH^r'Qi (4.72)
where Q2 is such that Qib'x' {I + H,Q2)b-,
[D-,\I + Q2H,),D-,'Q2]
(4.73)
is stable and det{I + Q2H1) 7^ 0 (det(I + H1Q2) ^ 0). The closed-loop eigenvalues are the poles of (4.73). (iv)
H2 = [D^'S2i]-'[D^'Q2]
(H2 =
[Q2b^'][Si2b^']-'),
(4.74)
where 5*21 (Su) and Q2 are such that Qib-,'
[D^'S2uD^'Q2] is stable with ^^^5*21 7^ 0 (detSu ^ 0) and ^'21 closed-loop eigenvalues are the poles of (4.75). (V)
H2 = [(I +
X2Ni)D^'r'X2
(4.75)
WVSnb-,\ Q2H2 = HSn-HiQ2
(H2 = X2[b-,\I
+ N,X2)r'),
= / ) . The (4.76)
where X2 (X2) is such that [(I +
X2Ni)D^\X2]
b^\l
X2 + NiX2)\
(4.77)
is stable and det(I + ^ 2 ^ ) 7^ 0 (det(I + N1X2) 7^ 0). The closed-loop eigenvalues are the poles of (4.77). Proof. The proof is straightforward in view of Theorem 4.8. Part (i) follows directly from Theorem 4.8. To show (ii), note that [Xi - KNi]Di - [-(Fi + KDiWi = I for any K (compare with Theorem 4.2). To show (iii), note that [Z)f ^7 + Q2Hi)]Di [D^^Q2]Ni = / for any Q2 (compare with Theorem 4.3). To show (iv), note that [D^^S2i]Di - [D^^Q2]Ni = / if and only if S21 - Q2H1 = 1. To show (v), note that [(/ + X2Ni)Dl^^]Di - [X2\Ni = I for any X2 (compare with Corollaries 4.6 and 4.7). What has not been shown yet is that all stabilizing H2 can be expressed by, say, (4.70) in (ii). This can be accomplished in a manner analogous to the proof of the theorem. In particular, any stabilizing H2 = 02^N2, where the pair (D2,iV2) is Ic, that gives rise to a closed-loop description (4.68) implies that bl^[D2, -N2] = [Xi - KNi, - ( F i + Xi Yi] KDi)] = [I,K] = [/, K]U, where U is a. unimodular matrix. Therefore, -Ni D i j from [/, K] = b}^^ 02, —N2W ^ it follows that a stable K can be determined uniquely. I Similarly, in (iii), the relation D ' ^ ^ i , A^2] = D^^[I + Q2H1, Q2] = D^^[I, Q] Hi determines uniquely a stable Q2 = Dibj^^N2' The details are left to the reader. These results were of course also shown in Theorems 4.2 and 4.3, using alternative approaches.
Remarks In Theorem 4.8 and Corollary 4.9, H\ may or may not be proper. Also, the stabilizing H2 may or may not be proper. Thus, the above results characterize all stabilizing H2, both proper and not proper. It is frequently desirable to restrict H\ and H2 to be proper rational matrices. The problem of interest then is to determine all proper stabilizing H2, given a proper Hi. Note that this problem has already been addressed previously in this section using the parameter Q2, and it will be studied at length in Subsection 7.4C. We conclude by summarizing the relations among the parameters used in this subsection. Note that the relations for all parameters, except Li, L2 and L2, L\, were derived in (4.58) to (4.64). The relations to L2, Li can easily be obtained in view of Corollary 4.9. These relations are summarized as:
and
L2
Xi-KNi
= D\\l
+ Q2Hx) = D^'S2i =(I + X2Ni)D^'
u
-{Y,+KD,)
L2
Xi - NyK = (1 + HiQ2)Di'
(4.78)
= D-,'Q2 = X2.
Also,
and
Li = -(Y,^D,K)
= Q2D-,'
EXAMPLE 4.6. Consider//i =
s+ 1
- SnDi'
= D^^I
+ N1X2)
=X2.
(4.79)
In view of Corollary 4.9, all stabilizing control-
lers are given by H2 = L2^L\, where the L2, L\ are stable and satisfy L2D1 -L\N\ = /. Now in view of (4.78) and (4.79), L2 = (/ + X2Ni)D\^ and Lx = X2. All appropriate X2, however, were characterized in Example 4.3 to be X2 = hid with d stable and d{l) + n{l)2 = 0. Therefore, all appropriate L2, L\ are given by L2 = d + h(s + I)
I
d
~
^ 1
,.
,1
i~.
1 ,
,
~^
; = -7, ^1 = -:, where d is stable and n is such that d + n{s + \) = d s- I d d (s- l)(i, i.e.,/2(1) = -^d(l). •
B. State Feedback and State Estimation State feedback and state estimation are studied in this subsection, using PMFDs of systems. Our current development, which offers additional insight, parallels that given in Chapter 4, where state-space descriptions were used. The study of state feedback and state estimation, using PMFDs, however, is important in its own right. An additional reason for discussing state feedback and state observers at this point is to provide the necessary background needed to connect the parameterizations of all stabilizing controllers using proper and stable factorizations with the internal descriptions (PMDs or PMFDs) of systems. This is accomplished in Subsection 7.4C. State feedback State feedback control laws are closely related to the state-space representations of systems. Recall from Chapter 4 that given X = Ax + Bu,
y = Cx -^ Du
(4.80)
605 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
606 Linear Systems
with A G /^"><^ B G /^"><^, C G iR^x^ and D G 7?^><^, the hnear state feedback control law is defined by u = Fx + Gr,
(4.81)
where F G /^^^^ and G G i^'^x^ with G nonsingular. Frequently, G - /. Then the closed-loop system description is given by i: = (A + BF)x + BGr,
y = {C + DF)x + DGr
(4.82)
and the closed-loop eigenvalues are the zeros of det [ql - (A + BF)]. In view of the relation
ql -A-BF - ( C + £>7^)
BG DG
^/ - ^ 5l \
-c
^
D\ [-F
^
G
(4.83)
it is not difficult to see that any geld of [ql - A, B\ will be an Id of ^/ - A - BF for any F. This implies that for complete eigenvalue assignment, ql — A and B must be left coprime, i.e., (A, B) must be completely controllable, which is a well-known result (see Chapter 4). Also, for stability, (A, 5) must be a stabilizable pair. Note that any geld of \_ql - A, 5] will be an Id of \ql - A- BF, BG] for any F and G. Since here G is taken to be nonsingular, [ql - A, B] and [ql - A- BF, BG] have the same geld, i.e., the open- and closed-loop systems have precisely the same uncontrollable eigenvalues. When G is singular, the closed-loop system may have additional uncontrollable modes (show this). Although F does not affect the uncontrollable modes of the system, it may alter its unobservable modes. For example, in an SISO controllable and observable system, it is possible to select F so that some of the closed-loop eigenvalues are at the same location as the finite zeros, and therefore, they become unobservable. This corresponds to pole/zero cancellations in the closed-loop transfer function. The closed-loop unobservable eigenvalues can also be studied by means of relation (4.83) and the gcrd of the pair (ql - A- BF,-C - DF), The effects of state feedback control laws can conveniently be studied using polynomial matrix descriptions. In particular, assume that the state-space representation (4.80) is controllable and in controller form [A^ Be, Cc, Dc) (see Chapter 3). An equivalent PMFD is then given by Dc(q)Zc(t) = u(t),
y(t) = Nc(q)Zc(t)
(4.84)
with Dciq) G R[q]'^'''^ and A^^(^) G R[q]P'''^, where Be Dc
0 ^PA
Dciq) -NAq)
Im 0.
ql - Ac -Cc
Be] \Sciq) Del [ 0
(4.85)
with the pair (B^ ql - Ac) being Ic and the pair (Ddq), Sdq)) being re [see (3.26)]. The matrix Sdq) = blockdiag[(1, q,..., ^^'~^)^] is an n X m matrix with J/, / = 1 , . . . , m, the controllability indices of {A^, Be, Ce, Dc). Note that the above relations can be written as {ql - Ac)Se{q) = BcDc(q),
Ndq) = CcSdq) + DcDdq),
(4.86)
which were derived via the Structure Theorem in Subsection 3.4D. Note that the states are related by xdt) = Sc(q)Zc(t). When the state feedback control law u(t) = FcXeit) + Gr{t) = FeSe(q)Zc(t) + Gr(t) - Fe(q)Zc(t) + Gr(t)
(4.87)
is applied, the closed-loop system state-space representation is {Ac + BcFc, BcG, Cc + DcFc, DcG} and the polynomial matrix description is Dp{q)Zc{t) = Gr(t\
y(t) = NF(q)zM
(4.88)
where Dj.(^) ^ Ddq) - Fdq) = : Dciq) - FcSc(q) and Npiq) = Ndql These representations are equivalent since Be Dc
0 \Dc(q) - FcSciq) -Nc(q) Ip\ [
G 0_
ql - Ac- BcFc ^c
J-^c^ c
BcG] \Sc(q) DCG\
0
[ 0 (4.89)
where the pair (Be, ql — Ac - BcFc) is Ic and the pair (Ddq) - FcSdq), Sdq)) is re (see Subsection 7.3 A). Note that (ql -Ac- BcFc)Sc(q) = Bc[Dc(q) - FcSc(q)] and NF(q) = (Cc + DcFc)Sc(q) + Dc[Dc(q)-FcSc(q)] = CcSc(q) +DcDc(q) = Nc(q\ i.e., the numerator Nc(q) is invariant under state feedback. Note that Nc(q) contains the zeros of the system (see Subsection 7.3B). Now assume that the state-space representation in (4.80) is controllable but not necessarily in controller form, and let A = Q~^AcQ, B = Q~^Bc, C = CcQ, D = Dc with Q a similarity transformation matrix. Relations (4.86) then assume the form (ql - A)S(q) = BDc(q),
Nc(q) = CS(q) + DDc(q\
(4.90)
where S(q) = Q-^Sc(q). Then Dc(q)Zc(t) = u(t\
y(t) = Nc(q)Zc(t)
(4.91)
is an equivalent polynomial matrix description and now x(t) = S(q)zc(t). Note that deg^. S(q) = deg^. Sc(q) = di, i = 1 , . . . , m, which are the controllability indices of the system. The linear state feedback control law is then given by u(t) = Fx(t) + Gr(t) = FS(q)zc(t) + Gr(t) = Fc(q)Zc(t) + Gr(tl
(4.92)
W e n o w h a v e ( ^ / - A - 5 F ) S ( ^ ) - B[Dc(q)-FS(q)\dindNc(q) = (C + DF)S(q) + D[Dc(q) - FS(q)] = CS(q) + DDc(q\ In view of the above, it can be seen that linear state feedback can be equivalently defined for the case of (controllable) polynomial matrix right fractional descriptions as shown in the following. Given D(q)z(t) - u(t),
y(t) = N(q)z(tl
(4.93)
where D(q) is column reduced, define the linear state-feedback control law by u(t) = F(q)z(t) + Gr(t), where deg^. F(q) < deg^. D(q) with F(q) closed-loop system is then described by DF(q)z(t) = Gr(t\
(4.94)
/^M^x^, G E 7?^x^, detG 7^ 0. The y(t) = NF(q)z(t),
(4.95)
where DF(q) = D(q) - F(q), NF(q) = N(q). The F(q) = FcSc(q) can be chosen to arbitrarily assign the polynomial entries of D(q), up to and including the terms of degrees J/ - 1 in the /th column of D(q), i ^ 1 , . . . , m. In fact, recall from the development in Chapter 3 (Theorem 4.10—the Structure Theorem) that D(q) — F(q) = D(q)-FcSc(q) = B'^ldiagiq^q-AmSc(q)]-FcSc(q) = B^^[diag[q^q-
607 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
608 Linear Systems
(Afn + BmFc)Sc{q)]. It was shown in Chapter 4 how to appropriately select Fc to arbitrarily assign all the closed-loop eigenvalues, i.e., the roots of det (D(q) - F{q)). Now consider the closed-loop state-space representation (4.82). As was discussed in Chapter 4, the closed-loop transfer function can be written as HF,G{^) = [(C + DF)[sI - (A + BF)]-^B + D]G = [C(sl - A)-^B + D][F[sI - (A + BF)Y^B
+ /]G
= H(s)He(s).
(4.96)
Here H(s) is the open-loop transfer function, and He(s) represents the transfer function of a system that, if connected in series with the given system, will apparently produce the same overall transfer function as the feedback system. Recall that this issue was addressed at length in Chapter 4. If the PMD (4.93) is used, then the closed-loop transfer function is given by =
HF,G(S)
NF(S)DF\S)G
=
N(s)D^\s)G
= [N(s)D-\s)][D(s)D^\s)]G
= H(s)He(s).
(4.97)
It is not difficult to verify that the system {Dpiq), Im, D{q), 0} is equivalent to the system {A + BF, B, F, Im}, both of which have the same transfer function He(s) (show this). Relation (4.97) also impHes that H(s) = N(s)D'\s)
=
HF,G(S)H;\S)
= [N(s)D^\s)G][D(s)D^\s)G]-\
(4.98)
Note that both HF,G and He are proper and stable (H(s) is proper). Furthermore, H~^ is also proper. Thus, HF,G and He are proper and stable factors in the MFD H(s) = HF,G(S)H;\S).
(4.99)
This is further discussed in the next subsection, where it is shown how all proper and stable right MFDs can be generated by means of state feedback. The left proper and stable MFDs can be generated from observers of the partial state. Observers of the state are examined next. State observers State observers were discussed at length in Chapter 4. Here we wish to present additional material concerning observers in terms of PMDs. Consider the plant S and the observer Sob of Fig. 7.6 and let the plant S be described by (4.93), where D(q) G R[qr'''^ and N(q) G Rlq^"^, As was discussed above, when D(q) is column reduced, the linear state feedback control law can be defined by u(t) = F(q)z(t) -h r(t), where deg^. F < deg^. D,i = 1 , . . . , m. Then the
u
^ +
1 *•
J '
1
H ^ob r w
y
o
'
FIGURE 7.6 Plant and observer
closed-loop system is represented by [D{q) - F{q)Ut)
609 y(t) == N(q)z(t).
= r(t\
(4.100)
When the state is not readily available, then a state observer for Fz may be used. Let the observer Sob be described by Q(q)Zob(t)=^
[K(qlH(q)]
u{t) (4.101)
W(0 = Zob(t),
where Q{q) E R[qT'''^. K(q) G Riq^'"'^, and H(q) G R[qr''P. Note that in Fig. 7.6 u(t) == w(t) + r(t). Assume now that the observer polynomial matrices Q, K, and H satisfy the relation K(q)D(q) + H(q)N(q) = Q(q)F(q),
(4.102)
where Q'^ and Q~^[K, H] are proper and stable (F G Rlq^'"'^). Recall that Q~^ proper is needed for Q(q)Zob(t) = 0 to be a "well-formed" set of differential equations so as to avoid impulsive behavior at ^ = 0. Note that if Q is, say, row or column proper, then Q~^ is proper. The rational function Q~^[K, H] is the transfer function of the observer. Then w = Zob = Q~^[Ku + Hy] = Q~^[KD + HN]z = Fz, i.e., Zob is a candidate for estimating a function F(q)z(t) of the partial state z(t). To show this, consider the closed-loop internal description D(q) -Q(q)F(q)
-Im \ z(t) • Q(q) [Zob(t).
Um r(t) [0_
(4.103)
),0] y(t) = [N(qlO] _Zol
Ku + Hy = (KD + 0 z(t) Im HN)z = QFz, and consider the unimodular transformation [-F(q) Zob(t). derived by using u = Dz = w + r = Zob -^ ^ and Qzob
z(t) Zob - F(q)z(t)
Z(t) . Then the closed-loop system description becomes e(t) D(q)-Fiq) 0
-I Z(t) Q(q)\ e(t)
r(t) (4.104)
zit) y(t) = [N(q\0] e{t)_ First, note that the closed-loop eigenvalues are the roots of detiP - F)detQ, and therefore, the closed-loop system is stable if and only if all the roots of both detiP - F) and detQ have negative real parts. The roots of det{D - F) are of course the closed-loop eigenvalues under state feedback when there is no observer, while the roots of det Q are the observer eigenvalues that are taken to be stable. Note that in view of Q(q)e(t) = 0, where Q~^(q) is proper and stable, the error ^(0 =" ZobO) - F(q)z(t) = w(t) - F(q)z(t) will go to zero as t goes to infinity, and therefore, the output w(t) of the observer will asymptotically approach the function of the state F(q)z(t). This will happen independently of r(t). Note that the roots of det Q are uncontrollable from r as can easily be seen, using, e.g., the eigenvalue test for
CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
610 Linear Systems
controllability. As expected, these uncontrollable eigenvalues cancel in the closedloop transfer function matrix given by
ms),o]
D(s)-F{s) 0
-1
--N{s)[D{s)-F{s)]-
(4.105)
In other words, after the transients caused by initial conditions have died out (see Chapter 4), the system behaves to the outside world as though an observer were not present. The observer in (4.101) was introduced in Wolovich [36]. For an observer (4.101) to exist, K,H, and Q must satisfy the Diophantine Equation KD + HN = QF given in (4.102), where QT^ and Q~^[K^H] are proper and stable to ensure causality. Note that if Dp = D — F, then F = D — Df and KD + HN = Q{D-DF), or {Q-K)D + {-H)N = QDp, which implies that [Q-\Q-K)][DD-p']
+
[-Q-'H][ND-p']=L
(4.106)
The pair ( g ^{Q — K)^—Q ^H) is a proper and stable solution of the equation XD + YN = /, where D = DD^^ and N = ND^^ are proper and stable. This equation is important in the parameterization of all stabilizing feedback controllers when using the ring of proper and stable rational functions, discussed in the next subsection. It is possible to implement the observer discussed above in an alternative manner. In particular, u = w^r = Q~^Ku^Q~^Hy + r impUes that ( / - Q~^K)u = QT^Hy + r,oxu = {I-Q-^K)-\Q-^Hy + r),ox u = {Q-K)-^Q{Q-^Hy
+ r).
(4.107)
This corresponds to the configuration in Fig. 7.7.
o
{Q-K)-^Q
Q-'H
FIGURE 7.7 Observer-based controller implementation Note that the feedback path controller Q~^H is always stable, while the controller in the feedforward path is biproper (i.e., it along with its inverse is proper) but not necessarily stable. If the external input r is of no interest, then take r = 0, in which case we have , u={Q-K)-^Hy. (4.108) This corresponds to the configuration depicted in Fig. 7.8.
cJ ^
"
S
y
ii
ff^ IX\ -1 U (f^ — T\j r7
FIGURE 7.8 Observer-based controller implementation when r --
It is of interest to compare these results with the corresponding state-space results in Chapter 4. In particular, consider the state-space plant representation (4.80) and assume that it is controllable and observable. Now comparing Fig. 7.7 with Fig. 4.7 of Chapter 4, we obtain, in view of (4.25) and (4.24) of Chapter 4, the relations (Q(s)-K(s))-'Q(s) = (I - Q-\s)K(s)r' = (I - Gu(s))-' = F[sI-{A-KC + BF - KDF)r\B - KD) -h /, Q-\s)H{s) = Gy(s) = F[sl - (A - KC)]-^K and Q-\s)K(s) = Guis) = F[sI-(A-KC)]-\B-KDXA\so,(Q(s)-K{s)r^H(s) = (/ - Gu(s))~^Gy(s) = F[sl -{A-KC-\-BFKDF)T^K (see Fig. 7.8). It is therefore clear that the two degrees of freedom controller in Fig. 4.7 of Chapter 4 is a special case (of order n) of the controller in Fig. 7.7. EXAMPLE 4.7. Consider ^(5*)
given in Example 4.2 of Chapter 4. In
view of the above, it is not difficult to see that the results developed in Example 4.2 can also be derived if we let 2(5) = s^ + d\s + d{),F{s) = (2-ai)s + (2-ao) = [^ao-l,22] = FS(sl K(s) s(2 - ai) - do(^ao - 1), and H{s) = s{{do - 2)(^ao ai 1) + (di - 2)(2 - ai)) + ((do di)(ao - 2) + (do - 2)(s - ai)X which satisfy (4.102). Verify this. • Finally, it remains to be shown that polynomial matrices K, H, and Q, which satisfy (4.102) with Q~^ and g " ^ [ ^ , / / ] proper and stable, exist. Here deg^.F < deg^. D, where D is assumed to be column reduced and the N, D are assumed to be re. The system S is assumed to be controllable and observable. That such K, H, and Q exist will not be shown here. This is shown in Wolovich [36], where an algorithm is given that is based on the Eliminant Matrix ofD and N (see Subsection 7.2E, Theorem 2.13, and Lemma 2.14) to select an appropriate row reduced Q and to determine K and H. Note that (4.102) can also be solved by using other methodologies, such as polynomial matrix interpolation (see the Appendix).
C. Stabilizing Feedback Controllers Using Proper and Stable MFDs Now consider systems ^i and ^2 connected in a feedback configuration as shown in Fig. 7.5. Let system ^i be controllable and observable and let it be described by its transfer function matrix Hi. In Subsection 7.4A, all systems ^2 that internally stabilize the closed-loop feedback system were parametrically characterized. In that development Hi was not necessarily proper, and the stabilizing H2 as well as the closed-loop system transfer function were not necessarily proper either. Recall that a system is said to be internally stable when all its eigenvalues, which are the roots of its characteristic polynomial, have strictly negative real parts. Polynomial matrix descriptions that can easily handle the case of nonproper transfer functions were used to derive the results in Subsection 7.4A, and the case of proper Hi and H2 was handled by restricting the parameters used to characterize all stabilizing controllers. Here we concentrate exclusively on the case of proper Hi and parametrically characterize all proper H2 that internally stabilize the closed-loop system. For this, proper and stable MFDs of Hi and H2 are used. These are now described. Consider H(s) G R(sy^^ to be proper, i.e., \ims-^ooH(s) < 00, and write the MFD as H(s) =
N'(s)D'(sy
(4.109)
611 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
612 Linear Systems
where the N'(s) andD'(s) are proper and stable rational matrices that we denote here as A^'(^) ^ RH^^'^mdD'is) G /?H^>^"'; that is, they are matrices with elements in/?^oo, the set of all proper and stable rational functions with real coefficients. For instance, s- 1 s- 1 (s - 2)(s + 1) if H(s) then H(s) (s - 2)(s + 1) (s + 2)(s + 3) (s + 2)(s + 3) 1 " 5 - 1 1 Is-2' are examples of proper and stable MFDs. _(S+ 1)2 J [s + l_ A pair (N\ D') G RHo, is called right coprime (re) in RHoo if there exists a pair (X', Y') G RHoo such that X'D' + Y'N' = L
(4.110)
This is a Diophantine Equation over the ring of proper and stable rational functions. It is also called a Bezout identity. Let H = N'D'-^ and write (4.110) as X' + Y'H - WK Since the left-hand side is proper, D'~^ is also proper, i.e., in the MFD given by H = N'D'~^, where the pair (N', D') is re, D' is biproper (D' and D'~^ are both proper). Note that X'~^, where X' satisfies (4.110), does not necessarily exist. If, however, H is strictly proper [lim^^oo H{s) = 0], then lim^^oo X'{s) = lim5_>oo D'{sy^ is a nonzero real matrix, and in this case X'~^ exists and is proper, i.e., in this case X' is biproper. When the Diophantine Equation (4.110) is used to characterize all stabilizing controllers, it is often desirable to have solutions {X', Y'), where X' is biproper. This is always possible. Clearly, when H is strictly proper, this is automatically true, as was shown. When H is not strictly proper, however, care should be exercised in the selection of the solutions of (4.110). LEMMA 4.10. luoXH = N[D'\^ = A^2^'2^ be rc factorizations. Then U', where U',U'
(4.111)
RHoo
Proof, Given the two re factorizations, let f/' = D\ D2 and note that A^' 2 = ^^'2 ^ N[D\^D'2 = N[U'.NowX!^D'2 + Y!^N^ = (X!^D[ + Y!^N[)U' = /, from which f/'-^ X^D; + r^A^j, i.e., U'-^ E RHoo. Similarly, X[D[ + Y[N[ = I implies that U' E RHoo. m Remarks 1. A matrix U\ as given above, with U' and f/'~^ G RHoo is a unit in the ring RHoo (refer to the discussion on rings and modules in Subsection 7.2E). 2. If// is also stable, i.e., if H E RHoo, then H = HI~^ is an re factorization. If now H = N'D'~^ in any re proper and stable MFD, then in view of Lemma 4.11, / - D'U\ i.e., D' and D'-^ G RHoo. EXAMPLE 4.8. Let 7/
1 (s-2)(s-\- 1)
and D' are re since X'D' + Y'N' =
1 Us+ 1)2/U+ 1
5 + 1 /\5 + 1
+ (9)
= A^'D'-^HereA^'
s- 1 (S + 1)2
1. In view of
=
s-1 Lemma 4.10, all
U'
••
s-\
613 t/', where t/',t/'-i G/?//oo are also re
CHAPTER?:
Polynomial Matrix factorizations of//. Here, U\s) = a{s)/b{s), where a{s) and b{s) are prime Hurwitz Descriptions polynomials of the same degree n and n can be arbitrarily large, but finite. • and Matrix Fractional The above example illustrates that when N' ^D' ^ RHoo dire re in RHoo, they may Descriptions have common factors that include common poles and zeros; however, these possible of Systems common poles and zeros have to be stable. Note that any common right divisor G^ of an re pair N^,D^ G RHoo must satisfy G^ G^~^ G RHoo. Contrast this with the case of two re polynomial matrices. Analogous results hold for Ic proper and stable MFDs of H = b'~^N' with b'^N' ^ RHoo, where the Diophantine Equation b'X'+N'f
=1
(4.112)
is satisfied for some X, F G RHoo. In this case the result corresponding to Lemma 4.10 becomes [52,^/^1 U'[b\,N[] with tJ', tJ'-^ GRHoo whenH - -&-^ Ni=b'-^Nl are Ic MFDs. As in the polynomial case, doubly coprime factorizations in RHoo of a transfer function matrix Hi = N[D'-^ = b'-^N[, where D[,N[ G RHoo and b[,N[ G RHoo are important in obtaining parametric characterizations of all stabilizing controllers. Assume therefore that X[
u'u'-
Y[
D\ N[
-n XI
I 0
0 /
(4.113)
where U^ is unimodular in RHoo, i.Q.,U'^U'~^ G RHoo. Also, assume that X[^X[ have been selected so that det X[ ^ 0 and det X[ ^ 0. To see how such relations can be derived [compare with the discussion following (4.18) in Subsection 7.4A], assume that some X'^^Y^ and Y^^X'^ have been found that satisfy X ^ D ; XN[ : / and b[X'^ -N[n- I. Note then that -N[
D'
D[ N'
D'S',-?;, N'Sl
-K
I 0
0 /
(4.114)
whereS', 4 x'X-YXLet {X[,Y[) = {X',X) and ( Z ( , - ? / ) = {NS'.+X'.^iys', Fj) to obtain (4.113). It can be shown that matrices are indeed unimodular. Internal stability Consider now the feedback system in Fig. 7.5 and let Hi and H2 be the transfer function matrices of Si and ^2, respectively, that are assumed to be controllable and observable. Internal stability of a system can be defined in a variety of equivalent ways in terms of the internal description of the system. For example in this chapter, polynomial matrix internal descriptions were used, and the system was considered as internally stable when its eigenvalues were stable, i.e., they had strictly negative real parts. In Theorem 3.15 in Subsection 7.3C, it is shown that the closed-loop feedback system is internally stable if and only if the transfer function between
and
614 Linear Systems
and
or
n
have stable poles, i.e., if and only if the poles of
/ -Hi
-H2
I
I
or
/ -1-1 0 Hi , respectively, are stable. 0 Hi -Hi In this subsection we shall regard the feedback system to be internally stable when I
/ -Hi
-H2 I
1-1
^RH^,
(4.115)
i.e., when all the transfer function matrices in (4.115) are proper and stable. In this way, internal stability can be checked without necessarily involving internal descriptions of 5i and ^2. This approach to stability has advantages since it can be extended to systems other than linear time-invariant systems. THEOREM 4.11. Leti/i = A^;Z)V^ = D'i^N[ mdH2 = D'^^N^ = N^D'^^ be doubly coprime MFDs in RHoo. Then the closed-loop feedback system is internally stable if and only if D ^ D ; -N2N[ = U'
(4.116)
D[D2-N{Ni
(4.117)
or if and only if = U',
where U', U'~^ G RHa. and U', U''^ G RHo.. Proof, Consider//i = N[D'^\H2 = D'2 ^^2 ^nd assume that (4.116) is satisfied. We h a v e / = D'2^N2N[D\^ + 6 2 ^ ^ ' ^ T ^ which implies that / -Hi / -Hi
and
D'2U'
-H2 I
-D'^'N^ -N[D'-,'
-H2 I
D'l
u'-^b'2^ 0
U'-^D'2^N':
(4.118)
which is proper and stable, since both factors in the right-hand side are proper and stable. Therefore, if (4.116) is satisfied, the closed-loop feedback system is internally stable. Similarly, it can be shown that if Hi = b'\^N[, H2 = Nl^D'^^ and (4.117) is satisfied, then the closed-loop feedback system is internally stable. The converse will now be established, namely, if the feedback system is internally stable, then (4.116) is satisfied. The proof that (4.117) is also true is completely analogous. Let Hi = N[D']^^ be re and H2 = ^^'2 ^A^2 ^^ ^^ MFDs in RHo., and let D ^ D ; - N^N[ = U' with U' some matrix in RH^. Recalling (3.103) or (4.37), we have I -Hi
-H2 '-' I
~
r (I-H2H1)-' [Hi(I-H2Hir
I +
{I'H2Hi)-'H2 Hi(I-HiH2r^H:2 j
(4.119)
where the identities (/-^1/^2)"^ = I + (I ~ HiH2)-^HiH2and(I - HiH2r^HiH2 = Hi{I - H2Hi)~^H2 were used. It is not difficult to see [compare also with (3.113)] that / -Hi
-H2 I
-1
'D[U' -'D'2 D[U"^N^ N[U'-^D'2 I + N[U'-^N2. "0 0" U'-'[D'2,N^l = 0 /. + [N[\
(4.120)
Assume now that the system is internally stable, i.e.,
/
-H2
e RHoo. Then,
I since D[, N[ are re and D2, N2 are Ic, (7 Ms in RHo,. To see this, premultiply U'~^[D2,N2l which is in RHa., by [D^,-A^^] G RH^, and postmultiply
by I -N: e RHo.. These operations leave the matrix in RHoo. Therefore, RH^ COROLLARY 4.12. Let//i = N{D'l^ = D'f^TVJ be doubly coprime MFDs in/?//oo, i.e., (4.113) is satisfied. Then the closed-loop feedback system is internally stable if and only if H2 has an Ic MFD in RHa,, H2 = D'2^N'2, such that D'2D[ -N'2N[ = /,
(4.121)
or if and only if H2 has an re MFD in RHoo, H2 = A/^2^'2 ^' ^^^^ that D[D'2-N[N2 = I. Proof. The proof is straightforward, in view of Theorem 4.11 and Lemma 4.10.
(4.122) •
Parameterizations of all stabilizing controllers Several parameterizations of all proper stabilizing controllers are now considered. Parameter K' THEOREM 4.13. Let//i = N[D'\^ = D'f ^A^j be doubly coprime MFDs in/?//oo that satisfy (4.113). Then all H2 that internally stabilize the closed-loop feedback system are given by H2 = -(X[ - K'N[)~\Y[ + K'D[) = - ( ? ; + D[K')(X[ - N[K')-\
(4.123)
where K' G RHoo is such that {X[ - K'N[)~^ [or (Xi - N[K')~^] exists and is proper. Proof. It can be shown that all solutions of 62^1 ~ ^2A^i = I are given by [D'2,
-N^]
U.K']
x[ rr -N[ D[\
(4.124)
where K' E RHoo. The proof of this result is similar to the proof of the corresponding result for the polynomial matrix Diophantine Equation in Subsection 7.2E and is omitted (see also Subsection 7.4A). Similarly, all solutions of D[D2 - N[N2 = / are given by (4.125) where K' E RHoo. The result then follows directly from Corollary 4.12. The above theorem is a generalization of the Youla parameterization of Theorem 4.2 over the ring of proper and stable rational functions. Generalizations of the Youla parameterization over rings other than the polynomial ring were introduced in Desoer et al. [11]; for a detailed treatment see also Vidyasagar [34]. It is interesting to note that in view^ of (4.113), H2 in (4.123) can be written as follows. Assume that X^^ and Xf^ exist. Then
615 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
616
H2 = -(Y[ + X'^\l
Linear Systems
= -[Y[X\\X[ = -Y[X\'
- Y[N[)K')(Xi
-
N[K')-^
- N[K') + X\^K'](Xi - X'i'K'iXi
- N{K'r'
-
N[KT^
= H20 + H2a,
(4.126)
i.e., any stabilizing controller H2 can be viewed as the sum of an initial stabilizing controller H20 = -Y[X'^^ and an additional controller H2a that depends on ^ ' . When K' = 0, then H2a is zero.
D[
eigenvalues, are the poles of /
H2
U'~\D2,
-1
1-1
/ -Hi
In view of (4.120), the poles of
-H2\ I
, which are the closed-loop
N2I Similarly, since
I + N^U
I =
I 0 + 0 0
U'-\N[,D[l
(4.127)
where Hi = D ' f ' ^ i i^ ^^ ^"'l ^ 2 = A^2^'2 ' is re, and both are M F D s in RHo. with D[D'^ - N[NL, U' and U', f/'"' G /?//„, it follows that the eigenvalues of the N'2 closed-loop feedback system are the poles of U'~^[N[, D[]. It is now straightD' forward to prove the following result. LEMMA4.15. Given//i = N[D\^ = 5 ' f ^ ^ j , doubly coprime MFDs in 7?//oo, if ^2 is given by (4.123), the closed-loop eigenvalues are the poles of 0 x: Y[ K'[N[,D[] (4.128) [X[ - Y[] [/, K'] -N[ D[\[0 -I -K' [N[,D[] I
or of
x: [N[,D[]-
K'[N[,D[l
(4.129)
Proof. Note that D'2D[-N^N[ = (X[-K'N[)D[+(Y[+K'D[)Ni = X[D[+Y[Ni I = U', which in view of the above discussion, directly implies the lemma.
= •
Therefore, in view of L e m m a 4.15, when the parameterization for H2 of Theorem 4.13 is used, the closed-loop eigenvalues are in general the poles of
[-N[ D[
m
, of
, andof/i:'.
1 fs-l\-i Ir' 1 NID'J' = 1 i-l- lVs+ 1 s + a/ s + a D\ N[ with a > 0, which are doubly coprime factorizations. Note that (s + 5)(s + a) s-\-3 ^ + 5-1 s- 1 s+ 1 (s + 1)(^ + 2) 1 0 X'l Y[ ^ + 2 s + 2 -Y[ -N[ D[ N'l ^ 1 0 1 (s + 3)(^ + a) 1 1 x: s -\- a s -\- a-^ s+ 1 (s + l)(s + 2) If all stabilizing H2 are parametrically characterized by means of (4.123), then in view EXAMPLE 4.9. Let Hi =
of Lemma 4.15, the closed-loop eigenvalues are in general the poles of
that are at X[ Y[ -1, the poles of that are at - 2 and -a, and the poles of K'. Also, H2 in this -N[ D[
case is given by
H2 =
617 . + 5 ^ ^ , . - l s + al _ s_±2 s+3 1 K'\ s+2 s+a
(s + (5 + {s + {s +
5)(s + 1)(^ + 3)(5 + 1)(5 +
1) 2) a) 2)
1 1^' V^ + 1
CHAPTER 7 :
ITK
V5 +
where K' G RHoo. Note that it is possible to select
-m1
so that the poles of X[ and Y[ are ^iJ
those of - A^{ and D[. In the above example, this is the case when a = 2. This is also the case when these quantities are expressed in terms of a state-space realization of Hi, as is shown later in this subsection. It is always possible of course to select K' so as to minimize the number of the closed-loop eigenvalues. This corresponds to minimizing the McMillan degree (number of poles) of//2- This follows from the fact that any proper stabilizing controller can be expressed as (4.123) for appropriate K'. If for example H\ can be stabilized by means of a real static H2, then a ^ ' E RHo, exists for this to happen. In this case the number of closed-loop eigenvalues is equal to the number of poles of Hi. It is not easy, however, to find such K' unless for example X[ = I and Y[ is real, in which case ^ ' = 0 a n d H 2 = ^ j In view of Lemma 4.15, it is recommended that the number of poles in
^1
and in [-A^{, D['\ be taken to be the minimum possible, which is the number of poles in Hi. Also, the number of poles in [X[, Y[] should be taken to be low. This is accomplished in a systematic way in the following, using state-space descriptions. If the desired stabilizing H2 is known, then the appropriate K' that will modify the initial //20 ^ X''[^Y[,io yield the desired H2, can be calculated from (4.122) as
K' =
-{Y[+X[H2){D[-N[H2r\
(4.130)
EXAMPLE 4.10. In the above example 7/2 = -(/? + 1), Z? > 0 characterizes all static stabilizing H2. Then for a = 1, we have K' =
s+5 s+2 -bs-
5+3 (^+1) s+2
3/7 + 2 8+2
s+b s+ 1
5-1 5+ 1 1
b+l (s + l)(bs + 3 / 7 - 2 ) (s + 2)(s + b) '
which will yield the desired H2 = -{b -\- 1). The closed-loop eigenvalue is in this case at -b, as can easily be verified. Parameters Q2, X'2 Parameters other than K' can also be used to parametrically characterize all stabilizing H2. These parameters were introduced in Subsection 7.4A, and in the following these results are modified to accommodate the MFDs in RHoo. THEOREM 4.16. Let Hi = N[D\ = D':[^N[ be doubly coprime MFDs in RHoo that satisfy (4.113). Then all H2 that internally stabihze the closed-loop system are given by (i) 7/2 = (/ + Q2Hi)-'Q2 = [D'-,\I + Q2Hi)r'[D'-,'Q2-\ = [{I + X'2N'i)D'-,']-'X'^, (4.131)
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
618 Linear Systems
where Q2 is such that U\^\l + 22^1, G2] ^ ^^00 and (/ + QiHiY^ exists and is proper; or where Z2 is such that [(/ + X'^N'^~^U\^, X'^ E RHo. and (/ + X'jN'^~^ exists and is proper. Or by (ii)
//2 = Qiil^H^QiY' X'2[D'-,\l +
= [Q2D'-,'W
+
H,Q2)D'-,'r (4.132)
N[X'2)r'
where Q2 is such that
Qi D'l^ E RHoo and (/ + Hi ^2)"^ exists and is proper; or 1 + H1Q2
where X2 is such that
X', E RHo, and {I + N[X'2)~^ exists and is proper. VD'-,\I + N[X'2)
Proof, First, notice the similarity of the results in this theorem and in Theorem 4.3 and Corollary 4.6 in Subsection 7.4A. To verify (i) directly, note that if H2 = ^'2 ^^2' ^^ solutions of b2D[ - A^2^| = / are given by [b'2, N!2\ = [(/ + X!2N[)D'i\ X^] E RH^
(4.133)
where X2 E RHoo is a parameter that we set equal to N2- Then in view of Corollary 4.12, the result in (i) that involves X2 follows directly. Notice that (/ + X2N[)D']^^ is biproper, and therefore, H2 in (4.131) is proper. Now if Q2 = D[X2, then the results involving Q2 also follow. Note that Q2 E RHo,. Part (ii) can be verified in an analogous manner. Here Qi = X'2D[. • In view of Lemma 4.15 and the discussion preceding it, the next result follows readily. LEMMA 4.17. Let Hi = N[D'\^ = D'l^N[ be doubly coprime. If H2 is given by (4.131), then the closed-loop eigenvalues are the poles of [D'i\l
+
Q2Hi\D'i'Q2]
U + QiHu Qi] [(I + X^N[)D'i\X^l
(4.134)
If ^2 is given by (4.132), then the closed-loop eigenvalues are the poles of Qib'^'
[Nib[]
=
Qi [Hi, I] 1 + H1Q2 X'2 D'-,\I-VN[X'2).
[N[,b[l
(4.135)
Proof, The proof of this result is straightforward, in view of Lemma 4.15 and the discussion preceding it. • An interesting case is when Hi is stable, as the following corollary shows. COROLLARY 4.18. Let ^ 1 E RHo.. Then all H2 that internally stabilize the closedloop feedback system are given by H2 = (I - K'HiY^K'
= K'(I -
HiK')-\
(4.136)
where K' E RH^ such that (/ - K'Hi) ^ [or (/ - H^K') ^] exists and is proper. Furthermore, the closed-loop eigenvalues are the poles of \[IK']
I -Hi
0
(4.137)
01 \I 0' l\ [HX /.
I 0' for 0 /. the doubly coprime MFD/fi = Hil'^ = /-^//i. Then (4.123) of Theorem 4.13 reduces to expression (4.136) for H2. The closed-loop eigenvalues are then given by the poles of (4.137), in view of Lemma 4.15. •
Proof, Note that in this case (4.113) can be written as
/ -H,
Compare this corollary with Theorem 4.16, where H2 is expressed in terms of Q2. In this case it is clear that K' = -Q2. Note that K' E RH00 suffices to guarantee stability. MFDs and internal representations Consider^ = N'D'~^ = D'"^A^', a doubly coprime factorization in i^ii/oo, i.e., (4.113) is satisfied. It is possible to express all proper and stable matrices in (4.113) in terms of the matrices of a state-space realization of the transfer function matrix H{s). In particular, we have the following result. LEMMA 4.19. Let {A, 5, CyD} be a stabilizable and detectable realization of H{s), i.e., H{s) = C{sl - A)-^B + D, which is also denoted by H(s) =
and with C D (A, B) stabilizable and (A, C) detectable. Let F be a state feedback gain matrix such that all the eigenvalues of A -f- BF have negative real parts, and let K be an observer gain matrix such that all the eigenvalues of A - A'C have negative real parts. Define A-KC U'
X' -N'
y b'
B- KD K (4.138)
-F -C
and
U' =
D' N'
-r X'
-D
A + BF
B
F
/
C + DF
D
K (4.139)
/-I/ Then (4.113) holds and// = N'D'-'^ = /)'"iA^'are coprime factorizations of/^.
Proof. Relation (4.113) can be shown to be true by direct computation and is left to the reader to verify. Clearly, U\ U' G RHoo. That A^', D' and D', N' are coprime is a direct consequence of (4.113). That N'D' = D' ^N' = H can be shown by direct computation and is left to the reader. In view of Lemma 4.19, U' and U' Y'
U' =
E RHoo in (4.113) can be expressed as
-F [si - (A - KC)Y'[B -C
- KD, K] +
/ -D
Ol I\ (4.140)
andt/'
=
D' N'
-r X'
F -Ir [sI-(A-^BF)r'[B,K] C ^DF
+ (4.141)
619 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
620 Linear Systems
These formulas can be used as follows. A stabilizable and detectable realization {A, B, C, D] of H{s) is first determined, and appropriate F and K are found so that A + BF and A-KC have eigenvalues with negative real parts. Then U^ and t/'"^ are calculated from (4.140) and (4.141). Note that appropriate state feedback gain matrices F and observer gain matrices K can be determined, using the methods discussed in Chapter 4. The matrices F and K may be determined for example by solving appropriate optimal linear quadratic control and filtering problems. All proper stabilizing controllers//2 = A^2^'2^ ^ D'2 ^^2 ^f theplant/i/i are then characterized as in Theorem 4.13. It can now be shown, in view of Lemma 4.19, that all stabilizing controllers are described by k = {A + BF - K(C + DF))x -V Ky + {B - KD)ri u = Fx + ri,r2 = y-{C
+ DF)x - Dri, n = K'{q)r2,
(4.142)
which can be rewritten as k = Ax + Bu + K(y - (Cx + Du)) u = Fx + K'{q){y - {Cx + Du)\
(4.143)
Thus, every stabilizing controller is a combination of an asymptotic (full-state/fullorder) estimator or observer and a stabilizing state feedback, plus K'(q)r2 with r2 = y - (Cx + Du), the output "error" (see Fig. 7.9). Let H = N'D'-^
= D'~^N'
(4.144)
be coprime factorizations in RHoo, where N', D', N', D' G RHoo. Also, let H = ND-^ = D-^N
(4.145)
be polynomial matrix coprime factorizations. The relation of proper and stable MFDs of H(s) to internal PMDs of the system is established in the next result.
FIGURE 7.9 A state-space representation of all stabilizing controllers
THEOREM 4.20. (i) The pair {N',D') e RHoo defines an re factorization of H{s) as in (4.144) if and only if there exists a rational matrix n with n, n~^ stable and DYl biproper such that
621 CHAPTER?:
Polynomial Matrix n. Descriptions Furthermore, if only n is stable and DYl is proper, then (N^D^) is a right factorization and Matrix but it is not necessarily coprime. Fractional (ii) Similarly, the pair (N^,&) G RHoo defines an Ic factorization of H(s) in RHoo Descriptions as in (4.144) if and only if there exists a rational matrix fl with fl, fl~^ stable and YID of Systems biproper such that [N',D'] =U[N,D]. (4.147) Furthermore, if only n is stable and YID is proper, then {N\&) is a left factorization but it is not necessarily coprime. (4.146)
Proof, The proof of this result can be found in Antsaklis [4] and will not be repeated here. • An interesting implication of Theorem 4.20 is the following: let H = ND~^ be a right PMFD and let D be column proper; Dz = u,y = Nz is di controllable PMD. Following the development in Subsection 7.4B, define the linear state feedback control law by (4.94), as w = F{q)z + Gr, det G 7^ 0, where deg^.F{q) < dQg^.D{q). The closed-loop system is then Dpz = Gr^y = Nz given in (4.95), where Dp = D — F. Now in view of Theorem 4.20, the relation D^'G
(4.148)
defines re proper and stable factorizations of H, when {DJ^N} is detectable. If the pair (N^D) is re, then Yl = D^^G. In fact, it is shown in Antsaklis [4] that all re factorizations in RHoo may be obtained by means of (4.148), i.e., by using linear state feedback on controllable and detectable realizations of H. It is interesting to note that if {A,5,C,D} is an equivalent state-space representation to {DJ^N} with F (F, G) the corresponding state feedback gain matrices. [sI-{AC + DF BF)]-^BG
G. This expression is the same as (4.141), when G = 1. Simi-
lar results can be derived for the Ic factorizations {N^,&) that are related to state observers. As an application of (4.148), consider Theorem 4.8 and Corollary 4.12, where it is shown that if the given plant Hi = NiD^^ = N[D'^^ where Ni^Di are re polynomial matrices and A^(, D\ are proper and stable matrices, then H2 is a stabilizing controller if and only if it can be written as H2 = L^^Li, where L2D1 — LiNi = / (Theorem 4.8), or if and only if H2 can be written as H2 = ^2"^^/^' where D^,D[ A^(A^( = / (Corollary 4.12). Assuming that Di is column proper, then in view of (4.148), the last relation can be written as [(D^^G)5^]Di - [{Dp^G)N^]Ni /, and therefore, the relation between L2,Li and D21N2 ^^ gi^^i^ by G-^DF
\L2\ \LI\
(4.149)
622 Linear Systems ^
^-1 EXAMPLE 4.11. Let H(s) = 1-^^^. TV = A^^ , where N = s - I md D = . ^.. .. ^^ 2)(5 + 1) (s - 2)(^ + 1). (i) If n = , \ , , , then A^' = NU, D' = DU, and X'D' + Y'N' = ^-^ ^-^ + s- 1 L (s + 1> ® If n = ,
;t;
, , , then X'D' + F W = j ^ ' ^ ^ ^ ^j • ^ ^ ^ ^ ^
(5 + 1)2(5' + 3)
(5 + 1)(^ + 2)
(5 + 1)(^ + 3)
+
o(^ + 3) (s - 1)(^ + 2) ^ (^ + 2) (^ + 1)2(^ + 3) Notice that in both (i) and (ii), 11, H"^ are stable and DU is biproper as required by Theorem 4.20. • The X-approach Given the transfer function H(s), instead of obtaining proper and stable factorizations to characterize all proper stabilizing controllers via a Diophantine Equation over the ring of proper and stable rational functions, the transformation A = l/(s -\- a),a> 0, may be used. Then one works with a Diophantine Equation that involves polynomial matrices in A. This transformation maps the stable region in the s-plane into a "stable" region in the A-plane and the point ^ = oo to the point A = 0 (see Pernebo [29] for a detailed discussion). This approach corresponds to working with proper and stable factorizations of H,N', and D' with all the poles of A^' and D' at -a and requires only polynomial matrix manipulations [here H = (l/(^ + a)")/ in Theorem 4.20]. Recall that a rational R(s) is proper (there are no poles at s = ^) if and only if R[(l - Xa)/X] has no pole at A == 0. Therefore, for proper stabilizing controllers, solutions of appropriate polynomial Diophantine Equations are sought where the denominator Z)2(A) of the stabilizing controller H2 has no A factors in det D2W, i.e., ^ 2 ^ ' ^ ) has no poles at A == 0. Note that in this case, all the poles of the solutions of the corresponding proper and stable Diophantine Equation will also be at -a, and stabilizing controllers obtained by this method tend to assign multiple closed-loop eigenvalues at -a. As an illustration, consider H(s) = (s - l)/[(s - 2){s + 1 ) ] (see the example above) and let A - l/(^ + 1). Then H(X) = (I - 2A)A/(1 - 3A). The polynomial Diophantine Equation in A is solved to obtain Z D + M = (-6A + 1)(1 - 3A) + 9(1 - 2A)A = 1. The corresponding (proper and stable) Diophantine Equation is obtained if we let A = l/(^ + 1) in X, A Y, N. Then the Z', D', Y', N' of case (i) of the above example are derived. If the controller u = Cy + r with C{s) = -9(s + l)/(s - 5) is used, all three closed-loop eigenvalues will be at - 1 . [C(s) = C(A) = -X'^Y"^ with A = l/(s + 1).]
D. Two Degrees of Freedom Controllers Consider the two degrees of freedom controller Sc in the feedback configuration of Fig. 7.10. Here SH represents the system to be controlled and is described by its transfer function matrix H(s) so that y(s) = H(s)u(s),
(4.150)
The two degrees of freedom controller 5 c is described by its transfer function matrix C(s) in u(s) =
m C(s)
[Cy(s\
(4.151)
Cr(s)]
r(s). Since the controller Sc generates the input u to SH by processing independently y, the output of SH, and r, the external input, it is called a two degrees of freedom controller. In the following, we shall assume that / / is a proper transfer function and shall determine proper controller transfer functions C that internally stabilize the feedback system in Fig. 7.10. The restriction that H and C are proper may easily be removed, if so desired. Note that in the development in Subsection 7.4C we assumed proper transfer functions in the feedback loop, while the development in Subsection 7.4A applies to nonproper transfer functions as well.
r
u Sc
y
SH
* FIGURE 7.10 Two degrees of freedom controller Sc
Internal stability THEOREM 4.21. Given is the proper transfer function H of SH, and the proper transfer function C of Sc in (4.151), where det(I - CyH) T^ 0. The closed-loop system in Fig. 7.10 is internally stable if and only if (i) u = Cyy internally stabilizes the system y = Hit, (ii) Cr is such that the rational matrix M = {I
(4.152)
-CyHY^Cr
(w = Mr) satisfies D ^M = X, a stable rational matrix, where Cy satisfies (i) and H ND~^ is a right coprime polynomial matrix factorization. Proof, Consider controllable and observable PMDs for SH, given by Dz = u,
(4.153)
y = Nz,
and for Sc, given by DcZc = [Ny,Nr]
U =
Zc.
(4.154)
where the N, D are re and the Dc, [Ny, Nr] are Ic polynomial matrices. The closed-loop system is then described by {bcD - NyN)z = Nrr, y = Nz
(4.155)
and is internally stable if the roots of det Do, where Do = DcD - NyN, have strictly negative real parts. (Necessity) Assume that the closed-loop system is internally stable, i.e., D~^ is stable. Since Cy = D~^Ny is not necessarily an Ic polynomial factorization, write [Dc,Ny] = GdDcy,Ncy],'^^^reGLis2igcldofthQpaiY0c,Ny).ThenDcyD-NcyN = Gl^Do = bk, where D^ is a polynomial matrix with D^^ stable; note also that G^^ is stable. Hence, u = Cyy = D'^^Ncyy internally stabiUzes y = Hu = ND'^u, i.e., part (i) of the theorem is true. To show that (ii) is true, we write M = (I - CyH)~^Cr =
623 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
624 Linear Systems
Dbl^Dcy0-^Nr) = Db^^Gl^Nr = DX, where X = D'^Nr is a stable rational ma^^^- ^^^^ shows that (ii) is also necessary. (Sufficiency) Let C satisfy (i) and (ii) of the theorem. If C = D~^[Ny, Nf] is an Ic polynomial MFD and GL is a geld of the pair (5c, Ny) then [Dc, Ny] = GiVDcy, Ncy] is true for some Ic matrices Dcy and Ncy(Cy = D^^Ncy)- Because (i) is satisfied, DcyD NcyN = D;t, where D^Ms stable. Premultiplying by GL we obtain Dc/)-A^^A/^ = GLDJ,. Now if G^^ is stable, then 5 ~ ^ where Do = DcD - NyN = GiDk, will be stable since 6^1 is stable. To show this, write D'^M = D-\l - CyUy^Cr = D^^Dcy0~^Nr) = b^^Gl^Nr and note that this is stable, in view of (ii). Observe now that the GL, Nr are Ic; if they were not, then C = b~^[Ny, Nr] would not be a coprime factorization. In this case no unstable cancellations take place in D^^G^^Nr 0^^ is stable), and therefore, if D~^M is stable, then {GLDkY^ = b~^ is stable or the closed-loop system is internally stable. • Remarks (1) It is straightforward to show the same results, using proper and stable factorizations of H given by H = N'D'-\
(4.156)
where the pair (N', D') G RHoo and (A^', D') is re, and of C = &;'[N;,N^I
(4.157)
where the pair 0'^, [N!^, N;.]) G RHOO and (D^, [A^^, TV/]) is Ic. The proof is completely analogous and is left to the reader. The only change in the theorem will be in part (ii) which will now read: (ii) Cr is such that the rational matrix M = (I - CyH)~^Cr satisfies D'-^M = X' G RH^, where Cy satisfies (i) and H = N'D'-^ is an re MFD in RHoo. (2) Theorem 4.21 separates the role of Cy, the feedback part of C, from the role of Cr, in achieving internal stability. Clearly, if only feedback action is considered, then only part (i) of the theorem is of interest; and if open-loop control is desired, then Cy = 0 and (i) implies that for internal stability H must be stable and Cr = M must satisfy part (ii). In (ii) the parameter M = DX appears naturally and in (i) the way is open to use any desired feedback parameterizations. In view of Theorem 4.21, it is straightforward to parametrically characterize all internally stabilizing controllers C In the theorem it is clearly stated [Part (i)] that Cy must be a stabilizing controller. Therefore, any parametric characterization of the ones developed in the previous subsections can be used for Cy. Also, Cr is expressed in terms of D~^M = X (or D' M ^ X'). THEOREM 4.22. Given that }) = //w is proper with// = ND~^ = D"^A^ doubly coprime polynomial MFDs, all internally stabilizing proper controllers C in w = CK are given by: (i)
C = (1 + QHr^[Q,M] = [(/ + LN)D-^]-^[L,Xl
(4.158)
where Q = DL and M = DX are proper with L, X and D'^I + QH) = (/ + LN)D-^ stable, so that (/ + QH)~^ exists and is proper; or (ii)
C = (Xi - KN)-\-{X2
+ Kb), XI
(4.159)
where K and X are stable so that {X\ — KN\) ^ exists and C is proper. Also, X\ and X2 ' Xi X2\ ID -X2 I 0" are determined from UU with U unimodular. -N 0 / D\ [N ^1. lfH= N'D'-^ = D'-^N' are doubly coprime MFDs in RHoo, then all stabilizing proper C are given by (iii)
C = {X[-K'N')-\-{X!^
+ K'D'),X%
(4.160)
where K',X' e RH^o so that {X[ -K'N')'^ exists and is proper. Also, U'U'-^ = D' 0" X',] 7 \ ^1 -xn with U'U'-^ eRK -N' & N' 0 /
K
c = (/+e//)-i[e,M]:
(iv)
[(/ + L W ) Z ) ' - l ] - l [ L ^ r ]
where Q = D'L'.M = D'X' e RH^o with L^X' and D'-^{I + QH) RHoo so that {I + QH)~^ or {I + LlN')-^ exists and is proper.
(4.161) {I + L'N')D'-
e
Proof, In view of part (i) in Theorem 4.21, Cy of C = [Cy, Q] can be expressed in terms of any parameterization of all stabilizing feedback controllers developed in Subsections 7.4A and 7.4C. Cr can then be expressed as Cr = {I — CyH)M with Cy given from above and M = DX with X any stable rational matrix. If Cy = {I + QH)-^Q with D-^[I + QH,Q] stable (see Theorem 4.3) or Cy = [{I + LN)D-^]-^L with [{I + LN)D-\L] stable (see Corollary 4.6), part (i) of the theorem follows. Notice that Cr = (I- CyH)M = (/ + QH)-^M = [{I + LN)D-^]-^X since Q = DL and M = DX. Similarly, if Cy = -{Xi- KN)'^ {X2 + KD) with K stable (see Theorem 4.2), then C, = {Xi - KNy^D'^M = {Xi - KN)-^X and part (ii) of the theorem follows. For part (iii), express Cy in terms of K' G RHoo, as in Theorem 4.13, and use X' G RHoo in part (ii) of Theorem 4.21, as in remarks following (4.157). Part (iv) follows from Theorem 4.16 of Subsection 7.4C. • Response maps It is straightforward to express the maps between signals of interest of Fig. 7.10 in terms of the parameters in Theorem 4.22. For instance u = C
- [C-y,C^
CyHu+Crr, from which we have u = {I — CyH)~^Crr = Mr. (In the following we will use the symbols u^y^ r, etc., instead of w, j , r, etc., for convenience.) If expressions (iv) of Theorem 4.22 are used, then u = D'X'r,
and
y = Hu=N'D'-^D'X'r
= N'X'r,
(4.162)
in view of {I — CyH)~^ = D'{I+ L'N')D'~^. Similar results can be derived using the other parameterizations in Theorem 4.22. To determine expressions for other maps of interest in control systems, consider Fig. 7.11, where du and dy are assumed to be disturbances at the input and output of the plant H, respectively, and 77 denotes measurement noise. Then, u = [C3;,Cr]
\y + dy + vi\
du,
from which we have
{I-CyH)-^[Crr + Cydy+Cy^+du] diXidy = Hu = H{l-CyU)-^[Crr CyVi+du]. Then, in view of (4.161) in Theorem 4.22, we obtain u = D'X'r + D'Lldy + D'L'I] +D\I + L'N')D'-^du = Mr + Qdy +
Qri+Sidu
+ Cydy
(4.163)
625 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
626 Linear Systems
o^ FIGURE7.il Two degrees of freedom control configuration and
y = N'X'r + N'L'dy + N'L'r] + A^'(/ + L'N')D'~^du = Tr + (So - I)dy + HQT] + HSidu.
Notice that Q = (I - CyHy^Cy or 7]. Also,
(4.164)
= D'L' is the transfer function between u and dy
Si = (I - CyHY^ = D'(I + L'N')D / - I
I + QH
(4.165)
is the transfer function between u and du. The matrix St is called the input comparison sensitivity matrix. Notice also that yo = y -^ dy = Tr-\- Sody + HQrf + HSidu\ i.e., So = {I-
HCy)
-1 _
(4.166)
I^HQ
is the transfer function between yo and dy. The matrix So is called the output comparison sensitivity matrix. The sensitivity matrices St and So are important quantities in control design. Now (4.167)
HQ = So- N'L' = I
since//g = H(I-CyH)-^Cy HCyil-HCy)-^ = -I + il-HCyT^ = -J + So. .yL±) ^y L±^yyL ±±^yj where So and HQ are the transfer functions from yo to dy and T;, respectively. Equation (4.167) states that disturbance attenuation (or sensitivity reduction) and noise attenuation cannot occur over the same frequency range (show this). This is a fundamental limitation of the feedback loop and occurs also in two degrees of freedom control systems. Similarly, we note that Si -QH
(4.168)
= I.
We now summarize some of the relations discussed above: T = H(I - CyHY'Cr M = (I
-CyHy^Cr
Q = (I-CyHy^Cy
= HM = NX,
{y = Tr)
DX,
(u = Mr)
= DL,
(u = Qdy)
So = {I-HCyY^
= I + HQ,
{yo = Sody)
Si = (I-
= I + QH,
(u = Sidu).
CyHy^
The input-output maps attainable from r, using an internally stable two degrees of freedom configuration, can be characterized directly. In particular consider the
two maps described by
627 CHAPTER 7 :
(4.169) i.e., the command/output map T and the command/input map M. Let H = ND ^ be an re polynomial MFD. THEOREM 4.23. The Stable rational function matrices T and M are realizable with internal stability by means of a two degrees of freedom control configuration [that satisfies (4.169)] if and only if there exists stable X so that (4.170)
X
Proof, (Necessity) Assume that T and M in (4.169) are realizable with internal stability Then in view of Theorem 4.20, X = Z)"^M is stable. Also, 3; = Hu = (ND'^)(Mr) = NXr. (Sufficiency) Let (4.170) be satisfied. If X is stable then T and M are stable. Also, note that T = HM. We now show that in this case a controller configuration exists to implement these maps (see Fig. 7.12). Note that u = Mr + Cy(fr -\- y) = [Cy, M + CyT]
r-y]
, from which we obtain \_r\ u = (I + CyH)-\M + Cyf)r.
(4.171)
HM this relation implies that u = Now if M = M and T = T, then in view of T (I + CyH)-\l + CyH)Mr = Mr and y = Hu HMr = Tr. Furthermore, Cy is a stabilizing feedback controller, and the system is internally stable since t and M are stable. •
m.\
1
A 1
M^M
o-C^ ;o
r
u
1
1
y
L
FIGURE 7.12 Feedback realization of (T, M) Note that other internally stable controller configurations to attain these maps are possible. (The realization of both response maps T and M, instead of only T as in the case of the Model Matching Problem, makes the convenient formation in Theorem 4.24 possible. The realization of both T and M is sometimes referred to as the Total Synthesis Problem; see [7], [8] and the references therein.) The results of Theorem 4.23 can be expressed in terms of H = N'D'~^, re MFDs in RHoo. In particular, we have the following result. THEOREM4.24. T,M EL RHOO are realizable with internal stability by means of a two degrees of freedom control configuration [that satisfies (4.169)] if and only if there exists
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
628
X' E RH^ so that
Linear Systems
\X'.
(4.172)
Proof. The proof is completely analogous to the proof of Theorem 4.23 and is omitted. It is now clear that given any desirable response maps
r such that
N' X\ where X' G RHo^, the pair (T, M) can be realized with internal staD' bility by using for instance a controller (4.161), C = [(/ + L'N')D'-T^[L',X% where [(/ + L'N')D'~^, L'] G RH^^ and X' is given above, as can easily be verified. It is clear that there are many C that realize such T and M and they are all parameterized via the parameter L' E RHoo that, for internal stability, must satisfy the condition (/ + L'N')D'~^ E RHoo. Other parameterizations such as K' can also be used. In other words, the maps T, M can be realized by a variety of configurations, each with different feedback properties. Remark In a two degrees of freedom feedback control configuration, all admissible responses from r under condition of internal stability are characterized in terms of the parameters X (or M), while all response maps from disturbance and noise inputs that describe feedback properties of the system can be characterized in terms of parameters such as ^ or 2 or L. This is the fundamental property of two degrees of freedom control systems: it is possible to attain the response maps from r independently from feedback properties such as response to disturbances and sensitivity to plant parameter variations. EXAMPLE 4.12. We consider H(s) =
(s - m + 2)
and wish to characterize all (s - 2)2 proper and stable transfer functions T(s) that can be realized by means of some control configuration with internal stability. Let H(s) =
1 (s - 2f s + 2\ {s + 2)2
N'D"^
be an
re MFD in RH Then in view of Theorem 4.24, all such T must satisfy N' 'T = s+2 T ^ X' G RHoo. Therefore, any proper T with a zero at +1 can be realized via 1 a two degrees of freedom feedback controller with internal stability. Now if a single degree of freedom controller must be used, the class of realizable T(s) under internal stability is restricted. In particular, if the unity feedback configuration {/; Gff, 1} in Fig. 7.15 is used, then all proper and stable T that are realizable under [s - I' internal stability are again given by T N'X' X\ where X' = L' E RH^ s+2 s- 1 (s + 2)2 E RHoo, i.e., [see (4.184)] and in addition (/ + X W ) ^ ' ,1 ^ •+^^r , (s - 2)2 \5 + 2
^ (s-2fp(s) X' = Hx/dx is proper and stable and should also satisfy (s+2)dx + (s for some polynomial p(s). This illustrates the restrictions imposed by the unity feedback controller, as opposed to a two degrees of freedom controller. Notice that these additional restrictions are imposed because the given plant has unstable poles. • It is not difficult to prove the following result.
THEOREM4.25. T,M,SG RHOO are realizable with internal stability by a two degrees of freedom control configuration that satisfies (4.169) and (4.167) [S = S^ see Fig. 7.11 and (4.163), (4.164)] if and only if there exist X', L G RH^ so that
rri
0] \X''
[N'
= D'
M _S _
.0
0
+
where (/ + L'N')D'-^ G RHoo, Similarly, T, M, Q there exist X',L' G RHo, so that
where (/ + L'N')D'-^
0
(4.173)
./_
A^'J
[N' T' M = D' .0 Q\
ro]
RHoo are realizable if and only if
0 0 D'
(4.174)
G RH^o.
Proof, The proof is straightforward in view of Theorem 4.24. Note that S ox Q are selected in such a manner that the feedback loop has desirable feedback characteristics that are expressed in terms of these maps. •
Controller implementations The controller C = [Cy, Cr] may be implemented, for example, as a system Sc as shown in Fig. 7.10 and described by (4.154), or as shown in Fig. 7.12 with C = [Cy, M + CyT], where Cy stabilizes H and T, M are desired stable maps that relate r to j and r to u, i.e., y = Tr and u = Mr, There are also alternative ways of implementing a stabilizing controller C. In the following, the common control configuration of Fig. 7.13 is briefly discussed. It will be denoted by {R; Gff, Gf^}. The {R\ Gff, Gfb} Configuration Note that since U = [Cy, Cr]
- [GffGfb,
(4.175)
GffRIl
{R; Gff, Gfb) is a two degrees of freedom control configuration that is as general as the ones discussed before. To see this, let C = [Cy, Cr\ = D'c ^i^r ^rl be an Ic MFD in 7^7^00 and let R = N;., Gff = D
Gft =
N;.
(4.176)
Note that R and G/^ are always stable; also, G^j exists and is stable. Assume now that C was chosen so that (4.177)
D'D' - N'N' = U',
1
r
I1 1 1 1 1 1 L
R
PC+ J • > !
Gff
u
H
y
,
FIGURE 7.13 Gfb
Iwo degrees or Ireedom controller {R; Off, Gfb}
629 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
630 Linear Systems
where U', U' E RHa,. Then the system in Fig. 7.13 with R, Gff, and Gft given in (4.176) is internally stable, as is shown in the following. In view of Theorem 4.21, the feedback system is stable if and only if (i) Cy = GffGfb = D'c'^Ny internally stabilizes H, and (ii) X' = D'-^M = D'-\l GffGfbHy^GffR G RHoo. It can be shown that (4.177) implies (i). Note that any possible cancellation in the product GffGft>, between D'c~^ and Ny, will involve only stable poles; this can easily be shown, using (4.177). The cancelled stable poles will be uncontrollable or unobservable eigenvalues in the closedloop system. In addition X' = D'-\l - GffGfbHy^GffR = D'-^[D'(D'^D' N!^N'y^D'c]£>'c^N;. = U'^N'r E RHoo. Therefore, the control configuration {R;Gff, Gfb} = {Nl\D'-^,N'y} with (4.177) satisfied is internally stable. In view of this result and Theorem 4.22, Gff, Gft,, and R may also be selected as R = X\ Gff = [(/ +
1-1-1 L'N')D'-^]
Ub
L'
(4.178)
WiihX', L,{I + L'N')D'-^ E J?//oo and J^r(/ + L W ) ^ 0, w h e r e r = D'-^Mand L' = D'-^Q, Now if X' and L' satisfy (4.173) or (4.174) of Theorem 4.25, desired command responses T, M and disturbance responses S ox Q are achieved under internal stability. Note that Gff, Gfb, and R may also be selected as R = X', Gff = (X[ - K'N'y
-1
, Gfb = -(X^ + K'N'),
(4.179)
where X',K' G RHoo. We shall now briefly discuss some special cases of the {R; Gff, Gfb} control configuration, which are quite common in practice. Note that the configurations below are simpler; however, they restrict the choices of attainable response maps and so the flexibility offered to the control designer is reduced when using these configurations. / . The {/; Gff, Gfb} controller. In this case u = [Cy, Cr] [GffGfb,
Gff]
, that is. Cv
CrG r^fb-
(4.180)
In view of (4.161) given in Theorem 4.22, this implies that L' = X'G fb>
(4.181)
or that the choice for the parameters L' and X' is not completely independent as in the {R; Gff, Gfb} case. The L' and X' must of course satisfy L', X' and (/ + L'N')D'-^ E RHa,. In addition in this case L' and X' must be such that a proper solution Gfb of (4.181) exists and no unstable poles cancel in X'Gfb. Note that these poles will cancel in the product Gff Gfb and will lead to an unstable system. Since L' and X' are
o FIGURE 7.14 The {/; Off, Gfb} controller
both stable, we will require that (4.181) have a solution Gfb ^ RHoo. This implies that if for example X'~^ exists, then X' and L' must be such that X'~^L' G RHoo, i.e., X' and L' have the same unstable zeros and L' is "more proper" than X'. This provides some guidelines about the conditions that X' and L' must satisfy. Also, Gff = [(I -hL'N')D'-T
(4.182)
X'.
It should be noted that the state feedback law implemented by a dynamic observer can be represented as an {/; Gff, Gfb} controller (see Subsection 7.4B, Fig. 7.8). 2. The {/; Gff, 1} controller, A special case of (i) is the unity feedback control configuration. Here u = [Cy, Cr]
; I.e.,
(4.183)
^
Gff
J
'
u
y
H
L
FIGURE 7.15 The {/; G//, /} controller which in view of (4.161) implies that X' = L'.
(4.184)
In this case the responses between y or u and r (characterized by X') cannot be designed independently of feedback properties such as sensitivity (characterized by L'). This is a single degree of freedom controller and is used primarily to attain feedback control specifications. 3, The {R\Gff,I}
[Gff,
controller. Here w - [C^;, C,] Cr — CyR.
GffR]
; I.e.,
(4.185)
In view of (4.161) given in Theorem 4.22, this implies that (4.186)
X' = L'R.
The L' and X' must satisfy L', X', (I + L'N')D'-^ e RHoo. In addition they must be such that (4.186) has a solution R G RHoo. Note that R stable is necessary for internal
R
^
Gff
u
H
y
FIGURE 7.16 The {R; Off, 1} controller
631 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
632 Linear Systems
stability. The reader should refer to the discussion in (1) above for the implications of such assumptions on X' and L'. Also, Off = [{I +
L'N')D'-^Y^L'.
(4.187)
4, The {R\ I, Gft,} controller. In this case U = [Cy, Cr]
r
tr^
1
•?
R
^
= [Gfb,R]
(4.188)
y
H
%
FIGURE 7.17 The {R; /, Gft} controller For internal stability, R must be stable. In view of (4.161) given in Theorem 4.22, this implies the requirement [(/ + VN')D'~^]~'^X' G RHoo, in addition to L', X', (I + L'N')D~^ E RHoo, which imposes significant additional restrictions on L'. Here
[Gfb,R] = [(I +
L'N')D'-T'[L\X'l
(4.189)
5. The {1,1, Gfh} controller. This is a special case of (4), a single degree of freedom case, where R = I. Here R = I implies that
X' = (/ + L'N')D'-\
(4.190)
or that X' and L' must satisfy additionally the relation D'X' - L'N' = I,
(4.191)
a (skew) Diophantine Equation. This is in addition to the condition that L', X', (I + L'N')D-^ G RHoo.
1
~\
Ji
G*
U
H
y
FIGURE 7.18 The {/; /, G/b} controller
Control problems In control problems, design specifications typically include requirements for internal stability or pole placement, low sensitivity to parameter variations, disturbance attenuation, and noise reduction. Also, requirements such as model matching, diagonal decoupling, static decoupling, regulation, and tracking are included in the specifications.
Internal stability has of course been a central theme throughout the book, and in this section all stabilizing controllers were parameterized. Pole placement was studied in Chapter 4, using state feedback, and output feedback via the Diophantine Equation was addressed earlier in Chapter 7. Sensitivity and disturbance/noise reduction are treated by appropriately selecting the feedback controller Cy. Methodologies to accomplish these control goals, frequently in an optimal way, are developed in many control books. It should be noted that many important design approaches such as the Hoo-optimal control design method, are based on the parameterizations of all feedback stabilizing controllers discussed earlier. In particular an appropriate or optimal controller is selected by restricting the parameters used, so that additional control goals are accomplished optimally, while guaranteeing internal stability in the loop. Our development of the theory of two degrees of freedom controllers can be used directly to study model matching and decoupling, and a brief outline of this approach is now given. Note that this does not, by any means, constitute a complete treatment of these important control problems, but rather, an illustration of the methodologies introduced in this section. In the model matching problem, the transfer function of the plant H{s){y = Hu) and a desired transfer function T{s){y = Tr) are given and a transfer function M{s)(u = Mr) is sought so that T{s)=H{s)M{s).
(4.192)
Typically, H{s) is proper, and the proper and stable T{s) is to be obtained from H{s) using a controller under the condition of internal stability. Therefore, M{s) can in general not be implemented as an open-loop controller, but rather, as a two degrees of freedom controller. In view of Theorem 4.24 (or Theorem 4.23), if H = N'D'~^ is an re MFD in RHoo, then the pair {T,M) can be realized with internal stability if and only if there exists X^ G RHoo so that
X\ Note that an M that satisfies
(4.192) must first be selected (there may be an infinite number of solutions M). In the case when det H{s) 7^ 0, T can be realized with internal stability by means of a two degrees of freedom control configuration if and only if N^~^T =X' ^ RHoo (see Example 4.12). In this case M = D'X'. Now if the model matching is to be achieved by a more restricted control configuration, then additional conditions are imposed on T for this to happen, which are expressed in terms of X' (see for instance Example 4.12 for the case of the unity feedback configuration and Exercises 7.23, 7.26). In the problem of diagonal decoupling, T{s) in (4.192) is not completely specified, but is required to be diagonal, proper, and stable. In this problem the first input affects only the first output, the second input affects only the second output, and so forth. \f H{s)~^ exists, then diagonal decoupling under internal stability via a two degrees of freedom control configuration is possible if and only if •ni_
di N'-^T
=N'-^
X' G RHo.
(4.193)
Clyn J
where H = N'D'-^ is an re MFD in RH^. and T{s) = diag [ni{s)/di{s)],i = l,...,m. It is clear that if H{s) has only stable zeros, then no additional
633 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
634 Linear Systems
restrictions are imposed on T{s). Relation (4.193) implies restrictions on the zeros ^^ ^i^^) when H{s) has unstable zeros. It is straightforward to show that if diagonal decoupling is to be accomplished by means of more restricted control configurations, then additional restrictions will be imposed on T{s) via X'. (See Exercise 7.25 and Exercise 4.17, 4.20 of Chapter 4 for the case of diagonal decoupling via linear state feedback.) The problem of diagonal decoupling has a long and interesting history and a very rich literature. The original solution of the problem involved linear state feedback and state-space descriptions and is due to Falb and Wolovich [12]. For an approach involving PMDs and the transfer function matrix, see Williams and Antsaklis [35]. Other types of decoupling, such as block, dynamic, and static, are also treated there. A problem closely related to diagonal decoupling is that of the inverse of H{s). In this case, T{s) = I. There is also a very rich literature on this problem and the interested reader is encouraged to find out more about it. A starting point could be Williams and Antsaklis [35]. (See also Exercises 4.17, 4.18 in Chapter 4.) In the problem of static decoupling, T(s) E RHa, is square and also satisfies r(0) = A, a real nonsingular diagonal matrix. An example of such T(s)
r ^2 + 1
s(s^ + 2) ]
where d{s) is a Hurwitz polynomial. d{s) [s(s + 2) s^ ^3s+ I Note that if T(0) = A, then a step change in the first input r will affect only the first output in y at steady-state, and so forth. Here y = Tr = T(l/s) and lims^osT(l/s) = T(0) = A, which is diagonal and nonsingular. For this to happen, with internal stability when H(s) is nonsingular (see Theorem 4.24), we must have N' T = X' G RHoo, from which can be seen that static decoupling is possible if and only if H{s) does not have zeros at 5* = 0. If this is the case and if in addition H{s) is stable, static decoupling can be achieved with just a precompensation by a real gain matrix G, where G = H-\Qi)K. In this case T{s) = H{s)G = H{s)H~\0)A, from which T(0) = A.
7.5 SUMMARY In this chapter alternatives to state-space descriptions were introduced and used to further study the behavior of linear time-invariant systems and to study in depth structural properties of feedback control systems. In Part 1, the properties of systems described by Polynomial Matrix Descriptions (PMDs) were explored in Section 7.3 and background on polynomial matrices was provided in Section 7.2. The Diophantine Equation, which plays an important role in feedback systems, was studied at length in Subsection 7.2E. An in-depth study of the theory of parameterizations of all stabilizing controllers with emphasis on PMDs was undertaken in Part 2, Subsection 7.4A, and the parameterizations of all proper stabilizing controllers in terms of proper and stable Matrix Fraction Descriptions (MFDs) were derived in Subsection 7.4C. State feedback and state estimation using PMDs were studied in Subsection 7.4B. Finally, control systems with two degrees of freedom controllers were explored in Subsection 7.4D, with an emphasis on stability, parameterizations of all stabilizing controllers, and attainable response maps.
7.6 NOTES Two books that are original sources on the use of polynomial matrix descriptions in systems and control are Rosenbrock [30] and Wolovich [36]. In the former, what is now called Rosenbrock's Matrix is employed and relations to state-space descriptions are emphasized. In the latter, what are now called Polynomial Matrix Fractional Descriptions (PMFDs) are emphasized, and the relation to state space is accomplished primarily by using controller forms and the Structure Theorem, which was presented in Chapter 3. Good general sources for the polynomial matrix description approach also include the books by Vardulakis [33], Kailath [21], and Chen [10]. Basic references for the material on polynomial matrices discussed in Section 7.2 are the books by MacDuffee [25] and Gantmacher [14]. These books also include material on the Diophantine Equation discussed in Subsection 7.2E. Additional sources for the properties of polynomial matrices that we found useful include Wolovich [36], Vardulakis [33], and Kailath [21]. A key concept in our development of polynomial descriptions for the study of systems is the notion of equivalence of representations, discussed in Subsection 7.3A, since it establishes not only relations between polynomial descriptions but also between polynomial and state-space representations. Original sources for this include Rosenbrock [30] and Fuhrmann [13]. See also Pernebo [28] and the comments by Rosenbrock in [31 ] and [32], noting that the definition of equivalence given by Wolovich [36] was shown by Pernebo in [28] to be the same as strict system equivalence. Additional material on this topic can be found in Kailath [21], Wolovich [36], and Vardulakis [33]. A good source for the study of feedback systems using PMDs and MFDs is the book by Callier and Desoer [9]. The development of the properties of interconnected systems, addressed in Subsection 7.3C, which include controllability, observability, and stability of systems in parallel, in series, and in feedback configurations is based primarily on the approach taken in Antsaklis and Sain [8], Antsaklis [2] and [3], and Gonzalez and Antsaklis [18]. Parameterizations of all stabilizing controllers are of course very important in control theory today. Historically, their development appears to have evolved in the following manner (see also the historical remarks on the Diophantine Equation in Subsection 7.2E): Youla et al. [37] introduced the K parameterization (as in Theorem 4.2) in 1976 and used it in the Wiener-Hopf design of optimal controllers. This work is considered to be the seminal contribution in this area. The proofs of the results on the parameterizations in Youla et al. [37] involve transfer functions and their characteristic polynomials. Neither the Diophantine Equation nor PMDs of the system are used (explicitly). It should be recalled that in the middle 1970s most of the control results in the literature concerning MIMO systems involved state-space descriptions and a few employed transfer function matrices. The PMD descriptions of systems presented in the books by Rosenbrock [30] and Wolovich [36] were only beginning to make some impact. A version of the linear Diophantine Equation, namely, AX ^-YB = C polynomial in z~^ was used in control design by Kucera in work reported in 1974 and 1975. In that work, parameterizations of all stabilizing controllers were implicit, in the sense that the stabilizing controllers were expressed in terms of the general solution of the Diophantine Equation, which in turn can be described parametrically. Explicit parameterizations were reported in Kucera [23]. In Antsaklis
635 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
636 Linear Systems
[1] the doubly coprime MFDs were used with the polynomial Diophantine Equation, working over the ring of polynomials, to derive the results by Youla in an alternative way. In Desoer et al. [11] parameterizations K' of all stabilizing controllers using coprime MFDs in rings other than polynomial rings (including the ring of proper and stable rational functions) were derived. It should also be noted that proper and stable MFDs had apparently been used earlier by Vidyasagar. In Zames [38], a parameterization Q of all stabilizing controllers, but only for stable plants, was introduced and used in //oo-optimal control design. (Similar parameterizations were also used elsewhere, but apparently not to characterize all stabilizing controllers; for example, they were used in the design of the closed-loop transfer function in control systems and in sensitivity studies in the 50s and 60s, and also in the "internal model control" studies in chemical process control in the 80s.) A parameterization X of all stabilizing controllers (where X is closely related to the attainable response in an error feedback control system), valid for unstable plants as well, was introduced in Antsaklis and Sain [7]. Parameterizations involving proper and stable MFDs were further developed in the 80s in connection with optimal control design methodologies, such as ii/oo-optimal control, and connections to state-space approaches were derived. Two degrees of freedom controllers were also studied, and the limitations of the different control configurations became better understood. By now, MFDs and PMDs have become important system representations and their study is essential, if optimal control design methodologies are to be well understood. In Subsection 7.4A, the discussion of parameters N^, D^, and K (Theorems 4.1 and 4.2) follows Antsaklis [1]. The material for the parameter X2 (Corollary 4.6) follows Antsaklis and Sain [7], where X2 was introduced in connection with the error feedback control configuration. Q2 was used in Zames [38] for stable systems (Corollary 4.4); however. Theorem 4.3 is valid for unstable systems as well. The discussion of the parameters Sn and Li, L2 follows Antsaklis and Sain [8], and Antsaklis [3]. Subsection 7.4B is based on Wolovich [36] and Antsaklis [2] and [3]. The results on proper and stable MFDs in Subsection 7.4C and their use in parameterizing all proper stabilizing controllers are due to Desoer et al. [11]. The development in Subsection 7.4C was based on Vidyasagar [34], Antsaklis [3], Green and Limebeer [20], and Maciejowski [26]. The development of the relations between MFDs in RHoo and PMDs of a system follows AntsakHs [4], and Nett et al. [27]; see also Khargonekar and Sontag [22]. The A-approach was developed in Pernebo [29]. The material on two degrees of freedom controllers in Subsection 7.4D is based on Antsaklis [3], Antsaklis and Gonzalez [6], and Gonzalez and Antsaklis [16], [17], [18], [19]; a good source for this topic is also Vidyasagar [34]. Note that the main stability theorem (Theorem 4.21) first appeared in Antsaklis [3] and Antsaklis and Gonzalez [6]. For additional material on model matching and decoupling, consult Chen [10], Kailath [21], Falb and Wolovich [12], Williams and AntsakUs [35], and the extensive list of references therein. 7.7 REFERENCES 1. P. J. Antsaklis, "Some Relations Satisfied by Prime Polynomial Matrices and Their Role in Linear Multivariable System Theory," IEEE Trans, on Autom. Control, Vol. AC-24, No. 4, pp. 611-616, August 1979.
2. P. J. Antsaklis, Notes on: Polynomial Matrix Representation of Linear Control Systems, Pub. No. 80/17, Dept. of Elec. Engr., Imperial College, London, 1980. 3. P. J. Antsaklis, Lecture Notes of the graduate course. Feedback Systems, University of Notre Dame, Spring 1985. 4. P. J. Antsaklis, "Proper, Stable Transfer Matrix Factorizations and Internal System Descriptions," IEEE Trans, on Autom. Control, Vol. AC-31, No. 7, pp. 634-638, July 1986. 5. P. J. Antsaklis, "On the Order of the Compensator and the Closed-Loop Eigenvalues in the Fractional Approach to Design," Int. J. Control, Vol. 49, No. 3, pp. 929-936, 1989. 6. P. J. Antsaklis and O. R. Gonzalez, "Stability Parameterizations and Stable Hidden Modes in Two Degrees of Freedom Control Design," Proc. of the 25th Annual Allerton Conference on Communication, Control and Computing, pp. 546-555, Monticello, IL, Sept. 30-Oct. 2, 1987. 7. P. J. Antsaklis and M. K. Sain, "Unity Feedback Compensation of Unstable Plants," Proc. of the 20th IEEE Conf. on Decision and Control, pp. 305-308, San Diego, December 1981. 8. P. J. Antsaklis and M. K. Sain, "Feedback Controller Parameterizations: Finite Hidden Modes and Causality," in Multivariable Control: New Concepts and Tools, S. G. Tzafestas, ed., D. Reidel Pub., Dordrecht, Holland, 1984. 9. F. M. Callier and C. A. Desoer, Multivariable Feedback Systems, Springer-Verlag, New York, 1982. 10 C. T. Chen, Linear System Theory and Design, Holt, Rinehart and Winston, New York, 1984. 11 C. A. Desoer, R. W. Liu, J. Murray, and R. Saeks, "Feedback System Design: The Fractional Approach to Analysis and Synthesis," IEEE Trans, on Autom. Control, Vol. AC-25, pp. 399-412, June 1980. 12. P. L. Falb and W. A. Wolovich, "Decoupling in the Design of Multivariable Control Systems," IEEE Trans, on Autom. Control, Vol. AC-12, pp. 651-659, 1967. 13. P. A. Fuhrmann, "On Strict System Equivalence and Similarity," Int. J. Control, Vol. 25, pp. 5-10, 1977. 14. F. R. Gantmacher, The Theory of Matrices, Vol. 1 and 2, Chelsea, New York, 1959. 15. Z. Gao and P. J. Antsakhs, "On Stable Solutions of the One and Two Sided Model Matching Problems," IEEE Trans, on Autom. Control, Vol. 34, No. 9, pp. 978-982, September 1989. 16. O. R. Gonzalez and P. J. Antsaklis, "Implementations of Two Degrees of Freedom Controllers," Proc. of the 1989 American Control Conference, pp. 269-273, Pittsburgh, PA, June 21-23, 1989. 17 O. R. Gonzalez and P. J. Antsaklis, "Sensitivity Considerations in the Control of Generahzed Plants," IEEE Trans, on Autom. Control, Vol. 34, No. 8, pp. 885-888, August 1989. 18 O. R. Gonzalez and P. J. Antsaklis, "Hidden Modes of Two Degrees of Freedom Systems in Control Design," IEEE Trans, on Autom. Control, Vol. 35, No. 4, pp. 502-506, April 1990. 19 O. R. Gonzalez and P. J. Antsaklis, "Internal Models in Regulation, Stabilization and Tracking," Int. J. of Control, Vol. 53, No. 2, pp. 411-430, 1991. 20. M. Green and D. J. N. Limebeer, Linear Robust Control, Prentice Hall, Englewood Cliffs, NJ, 1995. 21. T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980. 22. P. Khargonekar and E. D. Sontag, "On the Relation Between Stable Matrix Fraction Factorizations and Regulable Realizations of Linear Systems Over Rings," IEEE Trans, on Autom. Control, Vol. AC-27, pp. 627-638, June 1982. 23. V Kucera, Discrete Linear Control, Wiley, New York, 1979.
637 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
638 Lh^^Systems
24. V. Kucera, "Diophantine Equations in Control—A Survey," Automatica, Vol. 29, pp. 1361-1375, 1993. 25. C. C. MacDuffee, The Theory of Matrices, Chelsea, New York, 1946. 26. J. M. Maciejowski, Multivariable Feedback Design, Addison-Wesley, Reading, MA, 1989. 27. C. N. Nett, C. A. Jacobson, and M. J. Balas, "A Connection Between State-Space and Doubly Coprime Fractional Representations," IEEE Trans, on Autom. Control, Vol. AC29, pp. 831-832, September 1984. 28. L. Pernebo, "Notes on Strict System Equivalence," Int. J. of Control, Vol. 25, pp. 21-38, 1977. 29. L. Pernebo, "An Algebraic Theory for the Design of Controllers for Linear Multivariable Systems," IEEE Trans, on Autom. Control, Vol. AC-26, pp. 171-194, February 1981. 30. H. H. Rosenbrock, State-Space and Multivariable Theory, Nelson, London, 1970. 31. H. H. Rosenbrock, "A Comment on Three Papers," Int. J. of Control, Vol. 25, pp. 1-3, 1977. 32. H. H. Rosenbrock, "The Transformation of Strict System Equivalence," Int. J. Control, Vol. 25, pp. 11-19, 1977. 33. A. L G. Vardulakis, Linear Multivariable Control, Wiley, New York, 1991. 34. M. Vidyasagar, Control System Synthesis. A Factorization Approach, MIT Press, Cambridge, MA, 1985. 35. T. Williams and P. J. Antsaklis, "Decoupling," The Control Handbook, Chap. 50, pp. 745-804, CRC Press and IEEE Press, Boca Raton, FL, 1996. 36. W. A. Wolovich, Linear Multivariable Systems, Springer-Verlag, New York, 1974. 37. D. C. Youla, H. A. Jabr, and J. J. Bongiorno, Jr., "Modern Wiener-Hopf Design of Optimal Controllers—Part II: The Multivariable Case," IEEE Trans, on Autom. Control, Vol. AC21, pp. 319-338, June 1976. 38. G. Zames, "Feedback and Optimal Sensitivity: Model Reference Transformations, Multiplicative Seminorms, and Approximate Inverses," IEEE Trans, on Autom. Control, Vol. 26, pp. 301-320, 1981.
7.8 EXERCISES 7.1. Given a polynomial matrix P{s) G R[sY^^ with rankP(s) = mm(p, m), write a computer program to reduce P(s) to a row proper (row reduced) matrix via row elementary operations and to derive the corresponding unimodular matrix UL(S). Use your computer algorithm to verify the results of Example 2.10. Hint: Apply the algorithm to [P(s), Ip] so that in UL(S)[P(S), Ip] = [P(s), UL(S)], P(S) is in row proper form and UL(S) is the appropriate unimodular matrix. 7.2. Given a polynomial matrix P(s) G RlsV^"^ with p > m, write a computer program to reduce P(s) to column Hermite form and to derive the corresponding unimodular matrix UL(S). Use your computer algorithm to verify the results of Example 2.13. Hint: Apply the algorithm to [P(s), Ip] so that in UL(S)[P{S), Ip] = [P(s), UL(S)1 P(S) is in Hermite form and UL(S) is the appropriate unimodular matrix. 7.3. Consider A G R^^^ and let IAI//'"'!^ , be the r X r minor formed by selecting r rows I
\{ki,...,kr}
-^
^
denoted by {ii,..., ir}, and r columns of A, denoted by {^i,... Z:^}, where r < min(m, n). Let B G R'^^P and consider the minors of the product AB. These can be determined using \An\iH'-M
^ ^
I . |{'l.-.^r} |^|{/i,...,/r}
where {/i,..., Ir} denotes all possible sets of r integers among the n columns of A and n rows of 5, with r < mm{m,n,p). This formula for the minors of the product of matrices is known as the Binet-Cauchy formula. (a) \iA^R^^^,B^R^^^, with m m, write a computer program to determine a gcrd G^{s) of all the rows of P{s) and to derive the corresponding unimodular matrix U{s). Use your computer program to determine a gcrd and a geld of the matrices Pi{s) and P2{s) of Example 2.15. Hint: Determine a unimodular matrix U{s) such that U{s)P{s) that U{s)[P{s),Ip
G%{s) 0
0
. Apply your algorithm to [P(s), Ip] so
U{s)
7.5. Consider the polynomial matrices P{s) -
s^ + s -.2-1
R{s)-
s -s-l
(a) Are they re? If they are not, find a greatest common right divisor (gcrd). (b) Are they Ic? If they are not, find a geld. 7.6. (a) Show that two square and nonsingular polynomial matrices, the determinants of which are coprime polynomials, are both re and Ic. Hint: Assume they are not, say, re and then use the determinants of their gcrd to arrive at a contradiction, (b) Show that the opposite is not true, i.e., two re (Ic) polynomial matrices do not necessarily have determinants that are coprime polynomials. 7.7. (a) Show that if (the nonsingular) G^^ (s) and G^^ (s) are both gcrds of Pi and P2, then there exists a unimodular matrix UR{S) such that G^ (s) = UR{S)G^ (S). (b) Similarly, show that if the nonsingular G^ (s) and G|^ (s) are both gelds of Pi and P2, then there exists a unimodular matrix UL{S) such that G£ (5) = GI,{S)UL{S).
7.8. Show that v defined in (2.43) is indeed the observability index of the system. 7.9. Let P{s) = P„5" + P„_i5"-i + • • • +Po be a matrix polynomial with Pi e P"><" and let A G P"><". Show that (a) P{s) = Qr{s){sl-A) +Rr withRr = P„A" + P„_iA"-i + • •+Po, (b) P{s) = {sI-A)Qi{s)+Ri withRi =A"P„+A"-ip„_i + • •+Po. Hint: Qr{s) = PnS^'^ + (P„A + P„_i)5"-^ + • • • + {PnA^'^ + • • • + P i ) .
639 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
640 Linear Systems
7.10. Let P(s) be a polynomial matrix of full column rank and let y(s) be a given polynomial vector. Show that the equation P(s)x(s) = y(s) will have a polynomial solution x(s) for any y(s) if and only if the columns of P(s) are Ic, or equivalently, if and only if P(A) has full column rank for any complex number A. 7.11. Let P(s) = PdS^ + Pd-is"^'^ + • • • + Po ^ R[s]P'''^ and let Pe(s) = block diag (I^,..., Imy P(s)) with d blocks on the diagonal. It can be shown that by means of elementary row and column operations, Pe(s) can be transformed to a (linear) matrix pencil sE - A. In particular, Pe(s) = U(s)(sE - A)V(s), where E and A are real matrices given by E = block diag(Pd, Im, • •> Im)
A
-Pd-i /
•" •••
-Pi 0
-Po 0
0
•••
/
0
= •
and U(s) and V(s) are unimodular matrices given by I
s^-'l
si
U(s) =
y(s)
0 Bi(s)
si I
-I Bd-i(s)
with Bi+i(s) = sBi(s) + Prf-o+i), / - 1 , . . . , J - 2, and Bi(s) = sPd + Pd-i- Note that P(s) and sE - A have the same nonunity invariant polynomials. Let P(s) = s^ -\- s —s and obtain the equivalent matrix pencil sE - A. 1 7.12. Let {DR, I, NR} and {DL, NL, I] be minimal realizations of H(s), where H(s) •-^ NRDI' = DI^NL
X
with
Fl \DR
{NR,PR)
-Y X
re
and
(PL,NL)
Ic
and
related
by
I 0 where U is unimodular, i.e., they are doubly co0 /
NL DL\ [NR prime factorizations of H(s). Show that these realizations are equivalent representations. Hint: 0 NL 0 DR I DL NL NR -I -NR 0 0 0 / 0 / Y 0 / Y -X DR DL NL and also, X I I 0 -NR 0 0 /
7.13. Consider P{q)z{t) = Q{q)u{t) and y{t) = R(q)z(t) + W(q)u(t), where P{q) =
\q -q [-q-2
q 0
R(q) =
\2q^ + q + 2 -q-2
Q(q) = Iq 0
W(q) =
q- 1 1 1 1
-2q + 2 3q 3q + 4 -3q
with q = d/dt. (a) Is this system representation controllable? Is it observable? (b) Find the transfer function matrix//(>y)(y(5') = H(s)u(s)).
(c) Determine an equivalent state-space representation x = Ax + Bu, y = Cx + Du and repeat (a) and (b) for this representation. 7.14. Use system theoretic arguments to show that two polynomials d{s) = s"^ + dn-is""'^ + \-dis -\-do and n(s) = n„_i5'"~^ + nn-2S^~^ + \- nis -\- no art coprime if and only if Cc
rank
= n, CcAT'
0 0
1 0
0 0
0 1
where Ar =
a n d Q = [no, n\,..., 0 -do
0 -di
nn-\\.
1 -dn-\
0 -d2
s s+ 1
7.15. Consider the transfer function H{s)
Determine a minimal real-
ization in (a) Polynomial Matrix Fractional Description (PMFD) form, (b) State-Space Description (SSD) form. 7.16. In the following, assume that a PMD, [P, Q, R, W}, realizes an invertible mXm transfer function matrix H. (a) Show that the system matrix for H~^ is given by P -R
0
Q W
0
0
Im
Hint: Note thaiSdz^, -w^, - / ] ^ = [0,0, -u'^f when Pz = Qu,y = Rz + Wu. (b) Show that the following system matrix can also be used to characterize / / ~ \ R SM
P
— 0
-W Q
0
(c) Let detW ^ 0. Show that H'^ = RP'^Q -H W, where W = W'K Q = R = -W-^R, mdP = P + QW-^R.
QW'K
0 andA^ = [s^ -ls+ 1]. 0 (a) Determine an equivalent state-space representation. (b) Is the system controllable? Is it observable? Determine all uncontrollable and/or unobservable eigenvalues, if any. (c) Determine the invariant and the transmission zeros of the system.
7.17. Consider the system Dz = u, y = Nz, where D =
641 CHAPTER?:
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
642 Linear Systems
7.18. Consider the series connection depicted in Fig. 7.2 and the PMFDs for ^i and ^2, given in (3.63), (3.64) and (3.66), (3.67). (a) Show that the system S is controllable if and only if (i) (A^i, D2) is Ic, or (ii) (5iZ)2,M)islc, or (iii) (N2Ni,D2)is\c, (b) Determine similar conditions [as in (a)] for S to be observable. 7.19. Consider the parallel connection depicted in Fig. 7.1 and the PMFDs for ^i and S2 given in (3.52), (3,53) and (3.55), (3.66), Derive conditions for controllability and observability of S analogous to the ones derived for the series connection in Exercise 7.18. 7.20. Consider the double integrator//i = l/s'^. (a) Characterize all stabilizing controllers H2 for Hi using all the methods developed in Subsections 7.4A and 7.4C. (b) Characterize all proper stabilizing controllers H2 for Hi of order 1. 7.21. Consider the double integrator//i = l/s^. (a) Derive a minimal state-space realization for Hi and use Lemma 4.19 to derive doubly coprime factorizations in RHoo. (b) Use the polynomial Diophantine Equations (4.102) and (4.106) to derive factorizations in RHoo. 7.22. Consider//
s^ + I
s+1
(a) Derive a minimal state-space realization {A, B, C, D} and use Lemma 4.19 and Theorem 4.13 to parameterize all stabilizing controllers H2. (b) Derive a stabilizing controller H2 of order 3 by appropriately selecting K'. What are the closed-loop eigenvalues in this case? Comment on your results. Hint: A minimal state-space realization was derived in Example 3.2. The eigenvalues of A + BF and A - KC are stable, but otherwise arbitrary. Note that some of these eigenvalues become closed-loop eigenvalues. 7.23. Consider the unity feedback (error feedback) control system depicted in Fig. 7.19, where H and C are the transfer function matrices of the plant and controller, respectively.
a r
+ ^^
e
FIGURE 7.19 Assume that (/ + HC) ^ exists, (a) Verify the relations y = {1 + HCy^HCr + {I + HCT^d ^ Tr-\-Sd u = (1 + CHY^Cr - (/ + CHY^Cd = Mr - Md.
Compare these with relations (4.163) to (4.168) for the two degrees of freedom controller u = Cyy + G r . Note here that u = -Cy + Cr, and therefore, Cy = - C and Cr = C. Hence, for the error feedback system of Fig. 7.19, the relations following (4.168) assume the forms M = (/ + CHY^C
= DX = -Q =
T = H(I + CHY^C So =^ (1 + HCY^ Si = (/ + CHY^
(b)
-DL
= (/ + HCY^HC
= HM = NX
= I + HQ = I - HM =-
I-T
= I + QH = I - MH.
(i) Let H = ND~^ be an re polynomial factorization. Show that all stabilizing controllers are given by
C = [(/ - XN)D-^T^X, where [{I-XN)D~^, X] is stable and (I-XNY^ exists. Hint: Apply Theorem 4.22 to the error feedback case, (ii) If H is proper and H = N'D' is an re MFD in RH^, show that all proper stabilizing controllers are given by
C = [(/ - X'N')D' \'X\ where [(/ - X'N')D' v - l\ X'] G RH^ and (/ - X'N'Y^ exists and it is proper. Hint: Apply Theorem 4.22 to the error feedback case. (c) Assume that H is proper and H~^ exists, i.e., H is square and nonsingular. Let H = ND~^ be an re polynomial MFD. If T is the closed-loop transfer function between y and r, show that the system will be internally stable if and only if
[N-\I-T)H,N~^T] is stable. Assume that T ^ I for the loop to be well defined. Note that if T is proper, then
C = H-^T(I-TY^ is proper if and only if H~^T is proper and I - T is biproper. (d) Assume that in (c) H and T are SISO transfer functions. Let H = n/d. Show that the closed-loop system will be stable if and only if (1 - T)d
-1 _
Sd-
and
Tn-
are stable, i.e., if and only if the sensitivity matrix has as zeros all the unstable poles of the plant and the closed-loop transfer function has as zeros all the unstable zeros of the plant. Remark: Note that this is a result known in the classical control literature (refer to J. R. Ragazzini and G. F. Franklin, Sampled Data Control Systems, McGraw-Hill, 1958). It is derived here by specializing the more general MIMO case results to the SISO case. 5-1
(e) Given His) = ^rrz —, characterize all scalar proper transfer functions T that can be realized via the error feedback configuration shown in Fig. 7.19, under internal stability. For comparison purposes, characterize all T that can be realized via a two degrees of freedom controller and comment on your results. In both cases, comment on the location and number of the closed-loop eigenvalues. Hint: Use Theorems 4.22 and 4.24.
643 CHAPTER 7 :
Polynomial Matrix Descriptions and Matrix Fractional Descriptions of Systems
644 Linear Systems
(s - l)(s + 2)
7.24. Consider/^W
(s - 2)2
and characterize all proper and stable transfer func-
tions T(s) that can be realized via linear state feedback under internal stability. Hint: NDp^G and consider Theorem 4.23. From Subsection 7.4B, note that T 1 s+ 1 1
7.25. Consider// =
L^H- 1
5+ 3
1 5 + l-J
(a) Derive an re MFD in RHo., H = N'D'~ r^l
^1
(b) Let T
0
01 ni
and characterize all diagonal T that can be realized under in-
^2-i
ternal stability via a two degrees of freedom control configuration, (c) Repeat (b) for a unity feedback configuration {/; G//, /} (see Fig. 7.15). 7.26. In the model matching problem, the transfer function matrices H E RP^^{S) of the plant and T G RP'^'^is) of the model must be found so that T = HM [see (4.192)]. M is to be realized via a feedback control configuration under internal stability. Here we are interested in the model matching problem via linear state feedback. For this, let H = ND~^ an re polynomial factorization with D column reduced. Then Dz = u, y = Nz is a minimal realization of H. Let the state-feedback control law be defined by u = Fz + Gr, where F G Rlsr"""^, G G /?^><^ with detG y^O and deg^. F < deg^. D. To allow additional flexibility, let r = Kv,K ^ R"^""^. Note that (see Subsection 7.4C) HF,GK = NDp^GK = (ND-^)(DDp^GK) = {ND-^){D{G-^DF)-^K] = HM. In view of the above, solve the model matching problem via linear state feedback, determine F, G, and K, and comment on your results when (a) H
{s + \){s + 2) 2^2 - 3^ + 2 ' ^
(b) H
T
5+ 1 s+2
0 s
1 s s+2 s+ 1 (c) H = 1 L5+ 1
-^ + 2
T = h
s S + 3-]
s+2 0
s+ 1 5+ 4 -2 .(5 + 2)(5 + 4)J
//mr.' The model matching problem via linear state feedback is not difficult to solve when p = mmdrankH = m, in view of (G-^D^)"^/^ = D'^M = D'^H-^T = N'^T.
APPENDIX
Numerical Considerations
A.l INTRODUCTION To compute the rank of the controllabiHty matrix [B, AB,..., A^~^B\ or the eigenvalues of A, or the zeros of the system {A, B, C, D], typically requires use of a digital computer. When this is the case, one must deal with selection of an algorithm and interpret numerical results. In doing so, two issues arise that play important roles in numerical computations using a computer, namely, the numerical stability or instability of the computational method used, and how well or ill conditioned iht problem is numerically. An example of a problem that can be ill conditioned is the problem of calculating the roots of a polynomial, given its coefficients. This is so because for certain polynomials, small variations in the values of the coefficients, introduced, say, via round-off errors, can lead to great changes in the roots of the polynomial. That is to say, the roots of a polynomial can be very sensitive to changes in its coefficients. Note that ill conditioning is a property of the problem to be solved and depends on neither thefloating-pointsystem used in the computer nor on the particular algorithm being implemented. A computational method is numerically stable if it yields a solution that is near the true solution of a problem with slightly changed data. An example of a numerically unstable method to compute the roots of ax^ + Ibx + c = 0 is the formula {-b ± JQP- — ac))la that for certain parameters a, b, c may give erroneous results infinitearithmetic. This instability is due to subtraction of two approximately equal large numbers in the numerator when b^ » ac. Note that the roots may be calculated in a numerically stable way, using the mathematically equivalent, but numerically very different, expression c/(-b^ J(b^ - ac)). 645
646 APPENDIX:
Numerical Considerations
We would of course always like to use numerically stable methods and we would prefer to have well-conditioned problems. In the following section, we briefly discuss the problem of solving a set of algebraic equations given by Ax = b. We will show that a measure of how ill conditioned a given problem is, is the size of the condition number (to be defined) of the matrix A. There are many algorithms to numerically solve Ax = b, and we will briefly discuss numerically stable ones. Conditioning of a problem and numerical stability of a method are key issues in the area of numerical analysis. Our aim in this appendix is to make the reader aware that, depending on the problem, the numerical considerations in the calculation of a solution may be nontrivial. These issues are discussed at length in many textbooks on numerical analysis. Examples of good books in this area include, Golub and Van Loan [6] and Stewart [9], where matrix computations are emphasized. Also, see Petkov et al. [8] and Patel et al. [7] for computational methods with emphasis on system and control problems. For background on the theory of algorithms, on optimization algorithms, and their numerical properties, see Bazaran et al. [2] and Bertsekas and Tsitsiklis [3]. In Section 2 we present methods for solving linear algebraic equations. Singular values and singular-value decompositions are discussed in Section 3. In Section 4, an approach for solving polynomial matrix and rational matrix equations based on polynomial matrix interpolation is presented.
A.2 SOLVING LINEAR ALGEBRAIC EQUATIONS Consider the set of linear algebraic equations given by Ax = b,
(2.1)
where A E R^^^, Its solution is important in many engineering problems. It is of interest to know the effects of small variations of A and b to the solution x of this system of equations. Note that such variations may be introduced for example by rounding errors when calculating a solution or by noisy data. Condition number Let A E R^^^ be nonsingular. If A is known exactly and b has some uncertainty AZ?, associated with it, then A{x + Ax) = b + Afo. It can then be shown that the variation in the solution x is bounded by
llAxll
cond (A)
\m\
(2.2)
denotes any vector norm (and consistent matrix norm) and cond (A) dewhere | notes the condition number of A, where cond (A) ^ ||A||||A-i||. Note that cond{A)
(Tn
.(A)
.(AY
(2.3)
where crjnax(^), and o-mm(A) are the maximum and minimum singular values of A, respectively (see next section). From the property of matrix norms, ||AA~^|| <
||A||||A~^||, it follows that cond(A) > 1. This also follows from the expression involving singular values. If cond{A) is small, then A is said to be well conditioned with respect to the problem of solving Hnear equations. If cond{A) is large, then A is /// conditioned with respect to the problem of solving linear equations. In this case the relative uncertainty in the solution (||Ax||/||x||) can be many times the relative uncertainty in &(||AZ?||/||Z7||). This is of course undesirable. Similar results can be derived when variations in both b and A are considered, i.e., when b and A become Z? -h AZ? and A + AA. Note that the conditioning of A, and of the given problem, is independent of the algorithm used to determine a solution. The condition number of A provides a measure of the distance of A to the set of singular (reduced rank) matrices. In particular, if ||AA|| is the norm of the smallest perturbation AA such that A + AA is singular, and is denoted by d{A), then (iA/||A|| = l/cond (A). Thus, a large condition number indicates a short distance to a singularity, and it should not be surprising that this implies great sensitivity of the numerical solution xof Ax = bio variations in the problem data. The condition number of A plays a similar role in the case when A is not square. It can be determined in terms of the singular values of A defined in the next section. Computational methods The system of equations Ax = b is easily solved if A has some special form, (e.g., if it is diagonal or triangular). Using the method of Gaussian elimination, any nonsingular matrix A can be reduced to an upper triangular matrix U. These operations can be represented by premultiplication of A by a sequence of lower triangular matrices. It can then be shown that A can be represented as A = m
(2.4)
where L is a lower triangular matrix with all diagonal elements equal to 1 and U is an upper triangular matrix. The solution of Ax = b is then reduced to the solution of two systems of equations with triangular matrices, Ly = b and Ux = y. This method of solving Ax = bis based on the decomposition (2.4) of A, which is called the LU decomposition of A. If A is a symmetric positive definite matrix, then it may be represented as A = U^U,
(2.5)
where U is an upper triangular matrix. This is known as the Cholesky decomposition of a positive definite matrix. It can be obtained using a variant of Gaussian ehmination. Note that this method requires half of the operations necessary for Gaussian elimination on an arbitrary nonsingular matrix A, since A is symmetric. Now consider the system of equations Ax = b, where A G j^^nxn^ ^^^ j ^ ^ rank A = n(^ m). Then A = Q
= [Qh Qi\
= QiR.
(2.6)
where Q is an orthogonal matrix (Q^ = Q~^) and R G R^^^ is an upper triangular matrix of full rank n. Expression (2.6) is called the QR decomposition of A. When rank A = r, the QR decomposition of A is expressed as AP = Q
Ri 0
R2 0
(2.7)
647 APPENDIX:
Numerical Considerations
648 APPENDIX:
Numerical Considerations
where Q is orthogonal, Ri G R^^^ is nonsingular and upper triangular, and P is a permutation matrix that represents the moving of the columns during the reduction (in Q^AP). Q R decomposition can be used to determine solutions of Ax = b. In particular, consider A G R^^^ with rank A = n(< m) and assume that a solution exists. First, r DI
determine the 2/^ decomposition ofA given in (2.6). Then 2^Ax = Q^box\
X
=
0 Q^b (since Q^ = Q'^) or Rx = c. Solve this system of equations, where R is triangular and c = [In, 0]Q^b. In the general case when rank (A) = r < mm(n, m), determine the QR decomposition of A (2.7) and assume that a solution exists. The
\R^\c-R2y)]
c = [In OJG^Z?, where y G R"^'' is solutions are given by x = P\ y arbitrary. A related problem is the linear least squares problem where a solution x of the system of equations Ax = bis to be found that minimizes \\b - Ax\\2. This is a more general problem than simply solving Ax = b, since solving it provides the "best" solution in the above sense, even when an exact solution does not exist. The least squares problem is discussed further in the next section.
A.3 SINGULAR VALUES AND SINGULAR-VALUE DECOMPOSITION The singular values of a matrix and the Singular Value Decomposition Theorem play a significant role in a number of problems of interest in the area of systems and control, from the computation of solutions of linear systems of equations, to computations of the norm of transfer matrices at specified frequencies, to model reduction, and so forth. In the following we provide a brief description of some basic results and introduce some terminology. Consider A E %^^^ and let A* = A^, the complex conjugate transpose of A. A G ^"><^ is said to be Hermitian if A* = A. If A G /^^x^ then A* = A^, and if A = A^, thenA is ^jmm^mc. A G ^"^'^ is wn/^ryif A* = A~^ In this case A*A = AA"" = 4 . If A G /^'^xnhenA* = A^andif A^ = A'^, i.e., if A^A = AA^ = In, then A is orthogonal (refer to Section 6.2). Singular values Let A G ^^^^'^ and consider AA'' G ^ ^ x ^ . Let A/, / = 1 , . . . , m, denote the eigenvalues of AA'', and note that these are all real and nonnegative numbers. Assume that Ai > A2 ^ ---A^ > ••• > A;„. Note that if r === rank A === rankiAA''), then Ai > A2 ^ • • • ^ A^ > 0 and A^+i = • • • = A^ = 0. The singular values at of A are the positive square roots of A/, / = 1 , . . . , min(m, n). In fact, the nonzero singular values of A are A/2
CTi = (KY
/ = 1, .
> r.
(3.1)
where r = rank A, while the remaining (min(m, n) - r) of the singular values are zero. Note that cri > 0*2 ^ • • • ^ cr^ > 0, and cr^+i = (j^+2 = " ^ o-mm{m,n) = 0. The singular values could have also been found as the square roots of the eigenvalues of A* A G %''^'' (instead of AA* G ^"^x^). To see this, consider the following result.
LEMMA 3.1. Letm
649
n. Then AA*| = A'^-"|A/„ - A*A|,
(3.2)
i.e., all eigenvalues of A*A are eigenvalues of AA* which also has m - n additional eigenvalues at zero. Thus, AA* E ^'w><^ and A*A E ^"^^ have precisely the same r nonzero eigenvalues (r = rank A); their remaining eigenvalues, (m - r) for AA* and (n — r) for A*A, are all at zero. Therefore, either AA* or A*A can be used to determine the r nonzero singular values of A. All remaining singular values are zero. Proof of the lemma. The proof is based on Schur's formula for determinants. In particular, we have \X^'^J
A I
^''^ = \J^ A\ =
|Ai/^/^||Ai/2/,-A*A-i/^/^A| (3.3)
\X'^'aX-'^W\\In-A*A\
= A(^-«)/2 . |A/„ - A*A|, where Schur's formula was applied to the (1,1) block of the matrix. If it is applied to the (2, 2) block, then D(A) = A^«-^V2 . |A/^ - AA*|. Equating (3.3) and (3.4), we obtain |A/^ - AA*| = A^-"|A/„ EXAMPLE 3.1
|o
0
- Ar
= A.
0
(3.4) A*A|, which is (3.2). •
E R^''\ Here rank A = r = 1, A^AA*) = {5,0}, and Ai = 5, A2 = 0. Also, A/(A*A) =
"4 2 0" 2 1 0 .0 0 0_
and Ai = 5, A2 = 0, and A3 = 0. The
only nonzero singular value is ori = J~\i = + V5. The remaining singular values are zero. • There is an important relation between the singular values of A and its induced Hilbert or 2-norm, also called the spectral norm ||A||2 = \\A\\s, In particular, ,) = sup \\Ax\\2 = max{(AKA*A))i/2} = a(A), (3.5) M2 = l where &(A) denotes the largest singular value of A. Using the inequalities that are axiomatically true for induced norms (see Subsection I.IOB), it is possible to establish relations between singular values of various matrices that are useful in MIMO control design. The significance of the singular values of a gain matrix A(ja)) is discussed later in this section. There is an interesting relation between the eigenvalues and the singular values of a (square) matrix. Let A/, / = 1 , . . . , n, denote the eigenvalues of A E R^^^^ let A(A) = min/ |A/|, and let A(A) = max/ |A/|. Then l|A||2(-
a(A) < A(A) < A(A) < &(A).
(3.6)
Note that the ratio a(A)/a(A), i.e., the ratio of the largest and smallest singular values of A, is called the condition number of A, and is denoted by cond{A). This is a very useful measure of how well conditioned a system of linear algebraic equations
APPENDIX:
Numerical Considerations
650 APPENDIX:
Numerical Considerations
Ax = bis (refer to the discussion in previous section). The singular values provide ^ reliable way of determining how far a square matrix is from being singular, or a nonsquare matrix is from losing rank. This is accomplished by examining how close to zero a(A) is. In contrast, the eigenvalues of a square matrix are not a good indicator of how far the matrix is from being singular, and a typical example in the literature to illustrate this point is an n X n lower triangular matrix A with - I's on the diagonal and H-l's everywhere else. In this case, a(A) behaves as 1/2" and the matrix is nearly singular for large n while all of its eigenvalues are at - 1 . In fact, it can be shown that adding 1/2""^ to every element in the first column of A results in an exactly singular matrix (try it for n = 2). Singular-value decomposition Let A E ^^^^ with rank A = r < min(m, n). Let A* = A^, the complex conjugate transpose of A. THEOREM 3.1. There exist unitary matrices U G ^'"^'^ and V E ^"^"^ such that A = UXV\ ^r
•
where 2 =
^rX{n-r)
: LO(m-r)Xr
(3.7)
with Sr = diag(ai, (72,.. .,crr) E R^^^ se-
•
^(m-r)X(n-r)i
lected so that o-i > 0-2 ^ • • • > o"^ > 0. Proof. For the proof, see for example, Golub and Van Loan [6] and Patel et al. [7]. Let U = [Uu U2] with Ui E ^'"^^ U2 E ^^^(^-')
•
and V = [Vi, V2] with
Vi E ^'^^^^ V2 E ^"XC"-'-). Then A = f/SV* = UiXrVl
(3.8)
Since U and V are unitary, we have
u*u = and
v*v =
=Ir
(3.9)
[Vi, V2] = /„. VlVx = /,.
(3.10)
[Ul,U2] =Im,U\U, -y*-
Note that the columns of Ui and Vi determine orthonormal bases for 2/l(A) and 9l(A*), respectively. Now AA"" = (UiXrVDiVarUl)
IjT* = UiXjUl
(3.11)
from which we have AA*[/i = Ui^jUlUi
= UiXl
(3.12)
If Ui, i = 1 , . . . , r, is the ith column of C/i, i.e., Ui = [u\, U2,..., Ur], then AA*Ui = ajui,
i = l,...,r.
(3.13)
This shows thattheo"? are thernonzeroeigenvaluesofAA*,i.e.,CT/, / = 1 , . . . , r,are the nonzero singular values of A. Furthermore, M/, / = 1 , . . . , r, are the eigenvectors of AA* corresponding to cr?. They are the left singular vectors of A. Note that the ut are orthonormal vectors (in view of UIU\ = /r). Similarly, A'^A = (Vi^rUlXUiXrVl)
=
(3.14)
ViXjVl
from which we obtain A*AVi = ViS^Vt^i =
(3.15)
Vill
If v„ / = 1 , . . . , r, is the /th column of Vi, i.e., V\ = [vi, V2,. A*Avi = crjvi,
Vr], then
i = 1,2,..., r.
(3.16)
The vectors v/ are the eigenvectors of A*A corresponding to the eigenvalues a*?. They are the right singular vectors of A. Note that the vt are orthonormal vectors (in view
0fVlVi=Ir), The singular values are unique, while the singular vectors are not. To see this, consider Vi = VidiagieJ^O
and
f/i =
Uidiagie'^^^
Their columns are also singular vectors of A (show this). Note also that A = t/iSrVj implies that A =
^diUiv].
(3.17)
/=i
The significance of the singular values of a gain matrix A{jo)) is now briefly discussed. This is useful in the control theory of MIMO systems. Consider the relation between signals y and v, given by j = Av. Then max I * IHl2^o||v||2
or
= max MHb = ^ ( ^ ) IMb^o ||v||2
. m a x \\y\\2 = m a x ||Av||2 = &(A), IM|2 = 1 IH|2 = 1
(3.18)
Thus, cr(A) yields the maximum amplification, in energy terms (2-norm), when the transformation A operates on a signal v. Similarly, ,,min Ibib = min ||Av||2 = cr(A). IH|2 = 1
(3.19)
||v||2 = l
Therefore, QL(A)
llAvlb
IMb
(T(A),
(3.20)
where ||v||2 T^ 0. Thus the gain (energy amplification) is bounded from above and below by d-(A) and gi(A), respectively. The exact value depends on the direction of V.
651 APPENDIX:
Numerical Considerations
652 APPENDIX:
To determine the particular directions of vectors v for which these (max and min) gains are achieved, consider (3.17) and write
Numerical Considerations
Av = ^
(3.21)
aiUiv]v.
Notice that |v*v| < ||v/||||v|| = ||v||, since ||v/|| = 1, with equality holding only when V = avi,a E ^ . Therefore, to maximize, consider v along the singular value directions vt and let v = avi with |a| = 1 so that ||v|| = 1. Then in view of v]vj = 0, / # 7, and v]vj = I, i = j , we have that y = Av = aAvi = aaiUi and \y\i = \\Av\\2 = cTi, since ||w/||2 = 1. Thus, the maximum possible gain is cri, i.e., max ||j||2 = max ||Av||2 = ai(= (T(A)), as was shown above. This maximum gain hh = 1 llvlb = 1 occurs when v is along the right singular vector vi. Then Av - Av\ — a\U\ = y in view of (3.17), i.e., the projection is along the left singular vector u\, also of the same singular value a\. Similarly, for the minimum gain we have ar = ^(A) = min ||j||2 = min ||Av||2, in which case Av = Avr = (TfUr = y. IMb = 1
Iklb = 1
Additional interesting properties include gi(A) = 9l(f/i) = span{ui,...,
w^}
J{(A) = 91(^2) = span{vr+u • •., v j , where U = [MI, ..
Ur, Ur+h '",Um\
(3.22) (3.23)
= [Ui, U2] and V = [ v i , . . . , Vr, v^+1,..
Least squares problem Consider now the least squares problem where a solution x to the system of linear equations Ax = b is to be determined that minimizes \\b - Ax||2. Write min \\b - Ax\\l = mm(b - Axf{b - Ax) = mm(x^A^Ax - 2b^Ax + b'^b). Then X
X
X
Vjcix'^A'^Ax - Ib^Ax + b^b) = IPJAx ~ IPJb = 0 impUes that the x that minimizes \\b — Ax\\2 is a solution of A^AJC =
(3.24)
A^b.
Rewrite this as ViX^.Vjx = (UiXrViVb = Vi%Ulb in view of (3.14) and (3.8). Now X = ViX~^Ufb is a solution. To see this, substitute and note that VjVi = Ir. In view of the fact that X(A'^A) = J{(A) = ^(¥2) = span{vr+h • • •, v„}, the complete solution is given by x^ for some w E R"^~^. Since Vi^'^Ujb
ViX;^Ujb-^V2W
(3.25)
is orthogonal to V2W for all w,
xo = ViX^Ulb
(3.26)
is the optimal solution that minimizes |fc - Ax||2 (prove this). The Moore-Penrose pseudoinverse of A E R^^^ can be shown to be A* = ViS.-'f/ 1
•
(3.27)
It was seen that x = A ^Z? is the solution to the least squares problem. Furthermore, it can be shown that this pseudoinverse minimizes \\AA^ - Imllf, where \\A\\F denotes the Frobenius norm of A that is equal to the square root of trace[AA^] — YJ^^ihiAA^) = X r = i ^ ^ ( ^ ) - It is of interest to note that the Moore-Penrose pseudoinverse of A is defined as the unique matrix that satisfies the conditions (i) AA+A = A, (ii) A+AA+ = A+, (iii) (AA+f = AA+, and (iv) (A+Af = A+A. Note that if rank A = m ^ n, then it can be shown that A+ = A'^(AA^y^; this is in fact the right inverse of A, since A(A^(AA^y^) = Im- Similarly, if rank A = n^ m, then A+ = (A'^Ay^A'^, the left inverse of A, since ((A'^Ay^A'^)A = InSingular values and singular-value decomposition are discussed in a number of references. See for example Golub and Van Loan [6], Patel et al. [7], Petkov et al. [8],andDeCarlo[4].
A.4 SOLVING POLYNOMIAL AND RATIONAL MATRIX EQUATIONS USING INTERPOLATION METHODS Many system and control problems can be formulated in terms of matrix equations where polynomial or rational solutions with specific properties are of interest. It is known that equations involving just polynomials can be solved by either equating coefficients of equal power of the indeterminate s, or equivalently, by using the values obtained when appropriate values for s are substituted in the given polynomials. In the latter case one uses results from the classical theory of polynomial interpolation. Similarly, one may solve polynomial matrix equations using the theory of polynomial matrix interpolation. This approach has significant advantages. Full details can be found in AntsakHs and Gao [1] (see also Gao and AntsakHs [51). First, some required results from the theory of polynomial and rational matrix interpolation are briefly summarized. These are then used to determine solutions of polynomial and rational matrix equations. Polynomial matrix interpolation Consider first the polynomial case. The following is a fundamental result of the theory of polynomial interpolation: given / distinct complex scalars Sj, j = 1 , . . . , /, and / corresponding complex values bj, there exists a unique polynomial q(s) of degree / - 1 for which q(sj) = bj,
j = 1 , . . . , /.
(4.1)
Thus, an nth-degree polynomial ^(.s*) can be uniquely represented by the / = n 4- 1 interpolation (points or doublets or) pairs (sj, bj), j = 1 , . . . , /. The polynomial matrix interpolation theory deals with interpolation in the matrix case. In the following we cite a basic result upon which the solution to the polynomial matrix interpolation problem rests. Let S(s) = block diag([1, s,..., s^']^), where thedt, i = 1 , . . . , m, are nonnegative integers. Let aj ^ 0 and bj denote m X 1 and p X I complex vectors, respectively, and let Sj be complex scalars. THEOREM 4.1. Given interpolation triplets (sj, aj, bj), j = 1,..., I (i.e., interpolation points), and nonnegative integers dt with / = Sjl ^di + m such that the (Xjl^dt + m)xl
653 APPENDIX:
Numerical Considerations
654
matrix Si ^ [S(si)au...,S(si)ai]
APPENDIX:
Numerical Considerations
(4.2)
has full rank, there exists a unique p X m polynomial matrix Q(s), with /th-column degree equal to J/, / = 1,..., m, for which Q(sj)aj = bj,
j = h...J.
(4.3)
Proof. Since the column degrees of Q(s) are J/, Q(s) can be written as Q(s) = QS(s),
(4.4)
where the /? X (X/11 di + m) matrix 2 contains the coefficients of the polynomial entries. Substituting into (4.3), Q must satisfy QSi = Bu (4.5) where Bi = [b\, ...,bi\. Since Si is nonsingular, g, and therefore Q{s), are uniquely determined. • It should be noted that when/? = m =^ l a n d J i = l-l = n, the above theorem reduces to the Polynomial Interpolation Theorem. In that case, for Uj = 1, Si reduces to a Vandermonde Matrix that is nonsingular if and only if Sj, j = 1,... /, are distinct (show this). EXAMPLE 4.1. Let Q(s) b e a l x 2 = /?Xm polynomial matrix and let the following / = 3 interpolation points {(Sj, aj, bj), j = 1, 2, 3} be specified: {(-1, [1,0]^, 0), (0, [-1, 1]^, 0), (1, [0, 1]^, 1)}. In view of Theorem 4.1, Q(s) is uniquely specified when Ji and (^2 are chosen so that / = 3 = Xdi + m = {d\+d2) + 2,Qxd\+d2 = 1, assuming that ^3 has full rank. Clearly, there is more than one choice for d\ and ^2- The resulting Q{s) depends on the particular choice for the column degrees df. (i) LetJi = lmdd2 = 0.Then S(s) = block diag([hsf, I) and (4.5) becomes: ' 1 = Q[S(si)au S(s2)a2, S(s3)a3] = QS3 --Q - 1 0 = [0, 0, 1] = 53,
-1 0 1
0 0 1
from which we obtain g = [1, 1, 1] and Q{s) = QS(s) = [^ + 1, 1]. (ii) Let di = 0 and J2 = 1- Then S(s) = blockdiag(\, [hsf) and (4.5) yields Q = [0, 0, 1], from which we have Q(s) = [0, s], which is clearly different from (i). • Rational matrix interpolation Similar to the polynomial matrix case, the problem here is to represent a p X m rational matrix H(s) by interpolation triplets or points (sj, aj, bj), j = 1 , . . . , /, which satisfies Hisj)aj
= bj,
j = \,...,l,
(4.6)
where the Sj are complex scalars and the aj ¥= 0 and bj are complex m X 1 and p X 1 vectors, respectively. It can be shown that the rational matrix interpolation problem reduces to a special case of polynomial matrix interpolation. To see this, we write H{s) = D~^{s)N{s), where D{s) and N{s) are pX p and pX m polynomial matrices, respectively. Then (4.6) can be written as N(sj)aj = D{sj)bj, or as VN{sj), -D{sj)]
bj
= Q{sj)cj = 0,
;• =
l,...,l,
(4.7)
i.e., the rational matrix interpolation problem for a p X m rational matrix H(s) can be viewed as a polynomial interpolation problem for SL pX (p -\- m) polynomial matrix Q{s) = [N{s\ -D{s)\ with interpolation points {sj, Cj, 0) = (sj, [aj, b^jf, 0), j = 1 , . . . , /. There is also the additional constraint that D~^(s) exists. Solution of matrix equations In this segment, polynomial matrix equations of the form M(s)L(s) = Q(s) are considered. The main result is Theorem 4.2, which essentially states that all solutions M(s) of degree r can be derived by solving Eq. (4.16). In this way, all solutions of degree r of the polynomial equation, if they exist, are parameterized. The Diophantine Equation is an important special case and is examined at length. It is also shown that Theorem 4.2 can be applied to solve rational matrix equations of the form M(s)L(s) - Q(s), Consider the equation M{s)L(s) = Q(sl
(4.8)
where L(s) and Q(s) are given t X m and kx m polynomial matrices, respectively. We wish to determine thQ kX t polynomial matrix solutions M(s) when they exist. First consider the left-hand side of Eq. (4.8). Let M(s) = Mo + • • • + Mrs'
(4.9)
and let di = deg^^ [L{s)\, i = 1 , . . . , m, denote the column degrees of L{s), If Q{s) ^ M(s)L(s\
(4.10)
then deg^i [Q(s)] = di + r for i = 1 , . . . , m. According to the Polynomial Matrix Interpolation Theorem, Theorem 4.1, the matrix Q(s) can be uniquely specified, using ^T=\(^i + r) -h m = ^f=\di + m{r + 1) interpolation points. Therefore, consider / interpolation points {sj, Uj, bj), j = 1 , . . . , /, where Z - 2 r . i J / + m ( r + 1).
(4.11)
Let Sr(s) = block diag ([1, ^ , . . . , Z'+H^) and assume that the &Jlidi -h m(r +1)) X / matrix Sri =
[Sr(si)ai,...,Sr(si)ai]
(4.12)
has full rank, i.e., the assumptions in Theorem 4.1 are satisfied. Note that for distinct Sj, Sri will have full column rank for almost any set of nonzero aj. Now in view of Theorem 4.1, the matrix Q(s) that satisfies Q(sj)aj = bj,
;• =
h...J,
(4.13)
is uniquely specified, given these / interpolation points (sj, aj, bj). To solve (4.8), these interpolation points must be appropriately chosen so that the equation Q(s) ( = M(s)L(s)) = Q(s) is satisfied. We write (4.8) as MLris) = Q(s), (4.14) where M = [MQ, . . . , M^] and Uis) = [L'^(s),..., s''L'^(s)f are k X t(r + 1) and f(r + 1) X m matrices, respectively. Let s = sj and postmultiply by aj, j = 1 , . . . , /.
655 APPENDIX:
Numerical Considerations
656
Note that Sj and Uj, j = 1 , . . . , /, must be such that Sri has full rank. Define
APPENDIX:
Numerical Considerations
bj = Q{sj)aj,
j = 1,...,/,
(4.15)
and combine the above equations to obtain (4.16)
MLri = Bu where L^/ == [L^(^i)ai,..., L^(5'/)a/] and5/ = [^i,.. ,bi\2irtt{r+ matrices, respectively. THEOREM 4.2. Given L{s) and Q{s) in (4.8), let dt select r to satisfy degci [Q(s)] < di + r. Then a solution M(^) of (4.16) exists.
l)XlandkXl
degci [L{sy\, i = 1,..., m, and 1,
(4.17)
M[/, si,..., s^lY of degree r exists if and only if a solution M
It is not difficult to show that solving (4.16) is equivalent to solving ^i^j^j where
Cj = L(sj)aj,
= bp j = h. .,/, A j = 1,...,/. bj = Q(sj)aj,
(4.18) (4.19)
The M(s) that satisfy (4.18) are obtained by solving (4.20)
MSri = Bu
where Sri "= [Sr(si)ci,..., Sr(si)ci] has dimensions t(r + 1) X /, and Sr(s) [/, si, . . . , 5'''/]^ has dimensions t(r -\- I) X t, and Bi = [b\ , bi\ has dimensions kx I. Solving (4.20) is an alternative to solving (4.16). Constraints on solutions. When there are more unknowns than equations [t{r +1) and / = Sjl i<^/ + ^{r +1), respectively] in (4.16), this freedom can be exploited so that M{s) satisfies additional constraints. In particular, k = t(r+\)-l additional linear constraints, expressed in terms of the coefficients of M{s) (in M), can in general be satisfied. The equations describing the constraints can be used to augment Eqs. (4.16). In this case the equations to be solved become MiLruC] = [Bi,Dl
(4.21)
where MC = D represents the k linear constraints imposed on the coefficients M, and C and D are matrices (real or complex) with k columns each. • The Diophantine Equation An important case of (4.8) is the Diophantine Equation X(s)D(s) + Y(s)N(s) = Q(sl
(4.22)
where the polynomial matrices D(s), N{s), and Q(s) are given and X(s), Y(s) are to be determined. Note that if M(s) = [X(s\ Y(s)l
L{s) =
D(s) N(s)
(4.23)
then it is immediately clear that the Diophantine Equation is a polynomial equation of the form (4.8) and all previous results apply. Theorem 4.2 guarantees that
all solutions of (4.22) of degree r are determined by solving (4.16). In systems and control theory, the Diophantine Equation that is used involves a matrix L{s) = [D^{s), N^{s)Y, which has rather specific properties. These are exploited to solve the Diophantine Equation and to derive conditions for existence of solutions of (4.22) of degree r. It can be shown that the following result is true. THEOREM 4.3. Let r satisfy deg.i [Q{s)\ < di + r,
i = 1,..., m, and r •
1,
(4.24)
where v is the observability index of the system {D, /, N, 0}. Then the Diophantine Equation (4.22) has solutions of degree r that can be determined by solving (4.16) [or (4.20)]. EXAMPLE 4.2. Let D(s) = s-2 0
0 s+ 1
;v(^) =
s-\ 1
0 1
and
Qis)
1 0 0 1
We have di = di = 1, degd Q(s) = 0, / = 1, 2, and / = 2 + 2(r + 1). For r = I Sj = - 2 , - 1 , 0, 1, 2, 3, and 0 1 0 -1 -1 1 1 ' 3 ' -1 ' 3 ' 1 ' -1 A solution is given by M(s) = [X(s), Y(s)]
-S
5+1
0
-i.+i
Solving rational matrix equations Now let us consider the rational matrix equation (4.25)
M(s)L(s) = Q{s\
where L{s) and Q{s) are given t X m and k X m rational matrices, respectively. The polynomial matrix interpolation theory developed above can be used to solve this equation and determine the rational matrix solutions M(s) of dimension kx t. Let M(s) = D~^(s)N(s) be a polynomial fraction form of M(s) that is to be determined. Then (4.25) can be written as [A^(^), -D(s)]
L(s)
= 0.
(4.26)
= 0,
(4.27)
Note that one could equivalently solve [A^(^), -D(s)]
Lp(s)
where [Lp(s)^, Qp(s)^]^ = [L(s)^, Q(sW(t>(s) is a polynomial matrix with (l)(s) the least common denominator of all entries of L(s) and Q(s). In general, (l)(s) may be any matrix denominator in a right fractional representation of [L(s)^, QisW- The problem to be solved is now of the form (4.8), a polynomial matrix equation, where L(s) = [Lp(sf, Qp(sfV and Q(s) = 0. Therefore, all solutions [A^(^), -D(s)] of degree r can be determined by solving (4.16) or (4.20). Let s = Sj and postmultiply (4.27) by aj, j = 1 , . . . , /, with aj and / chosen properly. Define
657 APPENDIX:
Numerical Considerations
658
\Lp(s)]
APPENDIX:
[QP(S)\
Numerical Considerations
1,...,/.
(4.28)
The problem now is to determine a polynomial matrix [N(s), -D(s)] that satisfies [N(sj\ -D(sj)]cj
= 0,
J = 1 , . . . , /.
(4.29)
Note that restrictions on the solutions can easily be imposed to guarantee that D~^(s) exists and/or that M(s) = D~^(s)N(s) is proper. Additional constraints can be added so that the solution satisfies additional specifications [see (4.21)]. Pole placement Output feedback. All proper output controllers of degree r (of order mr) that assign all the closed-loop eigenvalues to arbitrary locations are characterized in a convenient way using interpolation results. Given N(s)D~^(s) = H(s), which is assumed to be proper, we are interested in solutions [X(s), Y{s)] of dimensions m X (p + m) of the Diophantine Equation where only the roots of \Q(s)\ are specified. Furthermore, X~^(s)Y(s) = C(s) should exist and be proper since it represents the controller. Here the equation to be solved is (X(sj)D(sj) + Y(sj)N(sj))aj
=0,
y = 1 , . . . , /,
(4.30)
orMLri = 0(1 = l.Jl^di + mr).Thus,thQ%^^^di + mrrootsof\X(s)D(s) + Y(s)N(s)\ are to be assigned the values Sj, j = 1 , . . . /. Note the difference between the problem studied earlier, where Q(s) is known, and the problem studied here, where only the roots of \Q(s)\ (or of |G('^)I within multiplication by some nonzero real scalar) are given. In the present case, the vectors aj can be viewed as design parameters and can be selected almost arbitrarily to satisfy requirements in addition to pole assignment. Note that this design approach is rather well known in the state feedback case, as is discussed later in this section (see also Chapter 4). The following result can be shown. THEOREM 4.4. Let r > v - 1. Then (X(s), Y(s)) exists such that all the n + mr zeros of \X{s)D{s) + Y{s)N{s)\ are arbitrarily assigned and X~^{s)Y{s) is proper. • 0 EXAMPLE 4.3. Let D{s) = Is-2 and A^(^) = s- 1 with n = L 0 s+l\ 1 deg \D(s)\ = 2. In this case there are deg \X(s)D(s) + F(5')A^(5)| = n + mr = 2 + 2r closed-loop poles to be assigned. Note that r > ^ ' - l = l - l = 0 . (i) For r = 0 and {(sp ajX j = 1, 2} = {(-1, [1, 0]^), (-2, [0,1]^)}, a solution of MLri = Ois M= -3 0 For this case, M = M(s) = [X(sX Y(s)] and C(s) = X-\s)Y(s) = ^ which 1 2 is a static output controller. (ii) For r = 1, and {(sj,aj)J = 1,...,4} = {(-1, [1,0]^), (-2, [0, If, (-3, -1,0]^), (-4, [0,-1]^)}, a solution of MLri = 0 is given by [X(s\Y(s)] = s-1 -1 12 5 + 1 ] . Note that C{s) = X{s) ^ Y{s) exists and is proper. 5+4 -6 5+4 State feedback. Let A, B, and F he n X n, n X m, and m X n real matrices, respectively. Note that 1^/ - (A + BF)\ = \sl - A\ • |/„ - {si - A)~^BF\ =
1^/ - A\ ' \lm — F(sl - Ay^B]. Now if the desired closed-loop eigenvalues Sj are different from the eigenvalues of A, then F will assign all n desired closed-loop eigenvalues Sj if and only if F[(sjl - Ay^Baj]
= Uj,
j = l...,n.
(4.31)
The m X 1 vectors aj are selected so that (sjl — A)~^Baj, j = 1,,. .,n, are linearly independent vectors. Alternatively, one could approach the problem as follows (see also Subsection 4.2B of Chapter 4). Let M(s) and D(s) be re polynomial matrices of dimensions nX m and mX m, respectively, such that (si - A)~^B = M(s)D~^(s). An internal representation equivalent to x = Ax -^ Bu in polynomial matrix form is Dz = u with X = Mz (see Subsection 7.3A of Chapter 7). The eigenvalue assignment problem now is to assign all the roots of \D(s) - FM(s)l or to determine F so that FM(sj)aj
= D{sj)aj,
J = h
(4.32)
Note that this formulation does not require that Sj be different from the eigenvalues of A as in (4.31). The m X 1 vectors aj are selected so that M(sj)aj, j = 1 , . . . , ^, are independent. Note that M(sj) has the same column rank as S(sj) = block diag([l, Sj,..., sj'~^]^), where dt are the controllability indices of (A, 5). Therefore, it is possible to select aj so that M(sj)aj, j = I,.. .,n, are independent, even when the Sj are repeated. In general there is great flexibility in selecting the nonzero vectors aj. For example when the Sj are distinct, which is a very common case, the aj can be selected almost arbitrarily. For all the appropriate choices of aj [M(sj)aj, j = 1,.. .,n, linearly independent], the n eigenvalues of the closed-loop system will be at the desired locations Sj, j = \,.. .,n. Different aj correspond to different F, which results in general in different closed-loop behavior. The exact relation of the eigenvectors to the aj can be determined by [sjl {A + BF)\M{sj)aj = (sjI-A)M(sj)aj-BFM(sj)aj = BD(sj)aj-BD(sj)aj = 0. Therefore, M(sj)aj = Vj are the closed-loop eigenvectors corresponding to Sj. One may select aj in (4.32) to impose constraints on the gains fij in F. For example, one may select aj so that a column of F is zero (take the corresponding row of all aj to be nonzero), or so that an element of F is zero. Alternatively, one may select aj so that additional design goals are attained. Note that this approach for eigenvalue/eigenvector assignment by state feedback has also been discussed in Subsection 4.2B of Chapter 4. For further details, see Antsaklis and Gao [1].
A.5 REFERENCES 1. P. J. Antsaklis and Z. Gao, "Polynomial and Rational Matrix Interpolation: Theory and Control Applications," Int. J. of Control, Vol. 58, No. 2, pp. 349-404, August 1993. 2. M. S. Bazaraa, H. D. Sherali, and C. M. Shetty, Nonlinear Programming Theory and Algorithms, 2d edition, Wiley, New York, 1993. 3. D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation—Numerical Methods, Prentice-Hall, Englewood Cliffs, NJ, 1989. 4. R. A. DeCarlo, Linear Systems. A State Variable Approach with Numerical Implementation, Prentice-Hall, Englewood CHffs, NJ, 1989.
659 APPENDIX:
Numerical Considerations
660 APPENDIX:
Numerical Considerations
5. Z. Gao and P. J. Antsaklis, "New Methods for Control System Design Using Matrix Interpolation," Proc. of the IEEE Conference on Decision and Control, Orlando, FL, December 1994. 6. G. H. Golub and C. F. Van Loan, Matrix Computations, The Johns Hopkins University Press, Baltimore, 1983. 7. R. V. Patel, A. J. Laub, and P. M. VanDooren, eds.. Numerical Linear Algebra Techniques for Systems and Control, IEEE Press, Piscataway, NJ, 1993. 8. P. Hr. Petkov, N. D. Christov, and M. M. Konstantinov, Computational Methods for Linear Control Systems, Prentice-Hall International Series, 1991. 9. G. W. Stewart, Introduction to Matrix Computations, Academic Press, New York, 1973.
INDEX
Abel's formula, 140 Ackermann's formula, 334, 353 Adjoint equation, 197 Adjoint of a matrix, 115, 442 Aircraft, 209, 320, 381 Algebra, 102 associative, 103 commutative, 103 Algebraic equation, 115 See also Solutions of algebraic equations Algebraic multiplicity, 124, 444 See also Eigenvalue Algebraic Riccati equation, 343, 349, 358, 363 See also Riccati equation Algebraically closed, 123 See also Field Ascoli-Arzela lemma, 22 Asymptotic behavior, 156, 186 See also Mode of system Asymptotic state estimator, 351 See also State observers Asymptotically stable, 159, 189, 449, 491, 494, 497, 504, 563 See also Equilibrium Asymptotically stable in the large, 450, 491, 494, 499 See also Equilibrium; Globally asymptotically stable At rest, 71,77 See also System Attractive, 449, 491 See also Equilibrium Automobile suspension system, 207, 380 Autonomous, 11, 62 See also Linear ordinary differential equation; System Axioms: of a field, 37 of a metric, 438 of a norm, 438 of a vector space, 38
Balanced representations, 430 Basis, 99, 100, 170, 439, 551, 552 See also Vector space Bessel's inequality, 440 Bezout identity, 553, 612 See also Diophantine equation BIBO stable, 481, 490, 505 See also Bounded-input/bounded-output stable Bijective, 101, 102 See also Linear transformation Bilinear functional, 435 signature, 435 symmetric, 435, 436 Binet-Cauchy formula, 639 Biproper, 552, 577 See also Rational function Block diagonal, 130 See also Matrix Blocking property, 307 See also Zero Bolzano-Weierstrass property, 18, 44 Bounded, 491, 493 See also Solutions of differential equations Bounded-input/bounded-output stability, 481,482,483,484,490,505 See also Stability Brunovsky canonical form, 288 See also Canonical form Building, earthquake protection, 208 B-W property, 18 See also Bolzano-Weierstrass property Canonical form, 116, 127, 535 Brunovsky, 288 Jordan, 130, 135, 136 Canonical Structure Theorem, 270 See also Kalman's Decomposition Theorem 661
662 INDEX
Cartesian product, 8 Cauchy criterion, 20 Cauchy sequence, 18 Cauchy-Peano existence theorem, 25 Causal, 67, 70, 71, 77 See also System Cayley-Hamilton Theorem, 124, 154 Characteristic equation, 123 Characteristic exponent, 163 Characteristic polynomial, 123, 299, 563 See also Matrix; Transfer function Characteristic value, 122 See also Eigenvalue Characteristic vector, 122 See also Eigenvector Chemical reaction process, 381 Cholesky decomposition, 647 Closed loop control, 327 Cofactor, 113 See also Matrix Column rank, 119 See also Matrix Column reduced, 526 See also Polynomial matrices Column vector, 107 See also Matrix Command input, 326 See also Input; Reference input Common divisor, 535 left, 536 right, 536 See also Divisor Commutative algebra, 103 See also Algebra Compact set, 9, 18 Companion form, 131, 153, 468 See also Matrix Complete instability, 473, 501 See also Equilibrium, unstable Complete space, 18 Completeness property, 288 Complex vector space, 38 Composite system, 203 See also System Condition number, 646, 647, 649 See also Matrix Conformal matrices, 109 See also Matrix Congruent matrices, 435, 436 See also Matrix Conservative system, 81 See also System Constructibility, 219, 247, 251, 252, 260 continuous-time system, 248, 252 discrete-time systeni, 219, 257 Gramian, 251,256, 262 See also Observability Continuation of solutions, 26, 27, 31,37 See also Solutions of differential equations Continuous dependence of solutions, 37 See also Solutions of differential equations Continuous function: over an interval, 9 at a point, 9, 44 uniformly, 9
Continuous time systems, 4 See also Systems Control problems, 632 ControllabiHty, 214, 215, 228, 235, 242, 264, 265, 274, 561 continuous-time system, 227, 235 discrete-time system, 215, 241 eigenvalue/eigenvector (PBH) test, 272 Gramian, 233, 240, 245 index, 283 matrix, 217 from the origin, 215, 217, 228, 236 See also Reachability totheorigin, 215, 218, 228 output, 312 subspace, 229, 236, 242 Controllable, 228, 235, 561 companion form, 279 differentially, 315 eigenvalues, 265 instantaneously, 315 mode, 265 pair, 229 single-input, 373 state, 242 subspace, 229, 236, 242 uniformly, 315 Controller: digital, 182 feedback, 589 implementations, 629 with two degrees of freedom, 368, 622 Controller companion form, 203 Controller form, 279, 280, 285, 291 multi-input, 283 single-input, 280 Convergence of a function, 18 pointwise, 18, 19,44 uniform, 18, 19,20 Convergence of a sequence, 18, 44 Convergence of a series, 21 pointwise, 21 uniform, 21 Converse theorem, 513 Converter (A/D, D/A), 183 Convolution: integral, 78 sum, 70 Coordinate representation, 100 See also Vector Coprime, 535, 593 See also Polynomial matrices; Polynomials Critical, 460, 462 See also Eigenvalue Cyclic, 132 See also Matrix D/A converter, 183 Dead-beat: control, 201 observer, 359 Decoupling: diagonal, 377, 378, 633 static, 634 Decoupling indices, 377 See also Indices
Degree, McMillan, 397 polynomial vector, 526 Detectable, 351 Determinant, 112, 121, 586 See also Linear transformation; Matrix; Return difference matrix Determinantal divisor, 299, 534 Diagonal decoupling, 377, 378, 633 Difference equations, 62, 63 See also Solutions of difference equations Differential equations, ordinary: classification, 11 first-order, 10 linear, 47 linear homogeneous, 13, 138 linear homogeneous with constant coefficients, 148 linear nonhomogeneous, 138, 145 nth order, 12 systems of first order, 37 See also Solutions of differential equations Digital controller, 182 Digital filter, 65 Digital to analog converter, 183 Dimension, 99 See also Vector space Diophantine equation, 540, 612, 656 all solutions, 550 Bezout identity, 553, 612 historical remarks, 552 Dirac delta distribution, 72, 74, 154 Direct form, 65 Direct link matrix, 169 Direct sum, 126 Discrete-time impulse, 68, 178 Discrete-time Kalman filter, 363 See also Kalman filter Discrete-time system, 3, 5, 60, 66, 174, 215, 241, 242, 257, 348, 358, 362, 388, 392, 401,489 See also System Distance function, 438 Divisor: common, 535 left, right, 536 greatest common, 535, 536 Domain of attraction, 449, 491 See also Equilibrium Double integrator, 205 Doubly coprime, 593 See also Coprime Dual system, 222, 402
Economic model, 205, 382 Eigenvalue, 121, 122, 298, 301, 305, 564 algebraic multiplicity, 124, 444 controllable, 265 critical, 460, 462 geometric multiplicity, 122 repeated, 129 spectrum, 122 Eigenvalue or pole assignment, 328, 330, 658 direct method, 328 eigenvector assignment, 337 using controller form, 332 Eigenvector, 121, 122
Elementary operations: column, 526 row, 526 Elementary unimodular matrices, 526 Eliminant matrix, 542, 547 e-approximate solution, 23 Equicontinuous, 22, 46 Equilibrium, 158, 189, 445, 446, 490 asymptotically stable, 158, 189, 449, 491, 494, 497, 504, 563 asymptotically stable in the large, 450, 491, 494, 499 attractive, 449, 491 completely unstable, 473, 501 domain of attraction, 449, 491 exponentially stable, 449, 477, 480, 512 exponentially stable in the large, 450, 459,471 globally asymptotically stable, 450, 491, 494, 499 isolated, 446 qualitative characterization, 447 stable, 159, 189, 448, 452, 455, 459, 491,492 trivial solution, 138, 446 uniformly asymptotically stable, 449 uniformly asymptotically stable in the large, 450,459,451 uniformly stable, 448, 453, 454, 455, 459, 461,470,512 unstable, 159, 189, 450, 459, 462, 472, 479, 481, 491, 495, 497, 499, 504, 513 See also Stability Equivalence: of internal representations, 170, 180, 554 relation, 535 transformation, 181 Equivalent, 116, 119, 172, 181, 535, 554 Estimator, 361 See also State estimator Euclidean: norm, 42 ring, 551 space, 437, 441 Euler: method, 25, 65, 85 polygon, 25 Exponentially stable, 449, 477, 480, 512 See also Equilibrium Exponentially stable in the large, 450, 459, 471 See also Equilibrium Extended system matrix, 557 External input, 326 External system description, 5, 65 See also System
Feedback, 326, 589 configuration, 203, 573 gain matrix, 327 integral, 379 output, 363, 379, 658 state, 321, 323, 324, 326, 327, 364, 658 Feedback control systems, 58, 589, 590 See also Feedback; Output feedback Feedback stabilizing controller, 589 paramererizations, 592
663 INDEX
664 INDEX
parameterizations, proper and stable (MFD), 611,615 two degrees of freedom, 622 Fermat's Last Theorem, 553 Field, 37 algebraically closed, 123 complex numbers, 38 rational functions, 37 real numbers, 38 Filtering theory, 357 Floquet multipliers, 163 Force field, inverse square law, 52 Fractional description, 521 See also System representations Frequency response, 204, 463, 465 Frobenius norm, 653 Function, 8 bijective, 101 continuous, 9 Hamiltonian, 81 indefinite, 437, 469 negative definite, semidefinite, 437, 491 one-to-one, or injective, 101 onto, or surjective, 101 piecewise continuous, 9, 58 positive definite, semidefinite, 437, 469, 491 Fundamental matrix, 139, 140 Fundamental sequence, 18 Fundamental set of solutions, 140 Fundamental theorem of linear equations, 101 Gap and position criterion, 465 See also Stability Gaussian elimination, 647 Generalized distance function, 468 See also Lyapunov function Generalized energy function, 468 See also Lyapunov function Generalized function, 74 See also Dirac delta distribution Geometric multiplicity, 122 See also Eigenvalue Globally asymptotically stable, 491, 494 See also Asymptotically stable; Equilibrium Gram determinant, 234 matrix, 234 Gramian constructibility, 251, 256, 262 controllability, 233, 240, 246 observability, 249, 253, 259 reachability, 230, 236, 245 Gram-Schmidt process, 440 Graphical criteria, 462 See also Stability Gronwall inequality, 29 Hamiltonian: dynamical system, 81 function, 81 matrix, 344, 349 Hankel matrix, 399 Hard disk, read/write head, 212, 382 Harmonic oscillator, 198 H-B property, 18,44 Heine-Borel property, 18, 44
Hermite form, 532 Hermitian matrix, 648 Highest degree coefficient matrix, 526 Homogeneous differential equations, 138 See also Differential equations Homogeneous solution, 58, 64, 145 See also Solutions of differential equations Hurwitz matrix, 460 Hurwitz polynomial, 462 Hybrid system, 3, 183 Idempotent matrix, 126 Identity matrix, 110 Identity transformation, 103 Ill-conditioned, 647 Impulse response: continuous-time, 75, 77 discrete-time, 70 time-invariant, 78 time-varying, 79 Impulse response matrix, 71, 77, 78, 79, 165, 166, 177 Index: bilinear functional, 437 nilpotent operator, 134 set, 17 Indices: controllability, 283 decoupling, 377 Kronecker, 288 observability, 295 Induced norm, 43 Infinite series, convergence, 21 Infinite series method, 150 See also Matrix, exponential Initial conditions, 4, 5 Initialtime, 4, 5, 61,62 Initial value problem, 10, 11, 12, 62 examples, 13 existence of solutions, 21 Inner product, 437 Inners, 498 Input: command, or reference, 326 comparison sensitivity matrix, 626 decoupling zeros, 302, 564 external, 326 function observability, 302 normal representations, 430 output decoupling zeros, 302, 564 output stability, 451,481 vector, 4 Input-output description, 5, 65, 165, 174 See also System representations Input-output stabiUty, 481, 505 See also Stability Instability, 159, 189, 450, 459, 472, 479, 481, 491,495,497,499,504,513 See also Equilibrium, unstable Integral equation, 11 Integral feedback, 379 Integral representation, 72, 75, 76 Integration, forward rectangular rule, 85 Interconnected systems, 568 feedback, 573 parallel, 568 series, 570
Interlacing, 464 See also Stability Internal description, 5 See also System representations Internal qualitative properties, 490 See also Internal stability Internal stability, 448, 451, 481, 490, 623, 624 Internally balanced realization, 424, 430 See also System representations Interpolation of polynomial, rational matrix, 653 Invariance principle, 493 See also Asymptotically stable Invariant factors, polynomials, 298, 534 Invariant property of (A, B), 288 Invariant subspace, 260 Invariant zeros, 302, 306, 563, 564, 565 See also Zero Inverse system, 377, 634 Inverse z-transformation, 178 Inverted pendulum, 89, 380 See also Pendulum Jacobian matrix, 48, 447 Jordan canonical form, 130, 135, 136 Kalman filter: continuous-time, 352 discrete-time, 363 Kalman-Bucy filter, 357 Kalman's Decomposition Theorem, 269,270 Kronecker indices, 288 Lagrange's equation, 83 Lagrangian, 83 Lamda approach, 622 Laplace transform, 75, 154 La Salle's Theorem, 493 See also Asymptotically stable; Invariance principle Latent value, 122 See also Eigenvalue Least squares, 648, 652 Least-order realization, 394 Left coprime, 536 See also Polynomial matrices Leonhard-Mikhailov stability criterion, 463 Level curve, 471 See also Lyapunov function Lienard equation, 16, 479 Limit cycle, 88 Linear algebraic equation: fundamental theorem, 101 homogeneous, nonhomogeneous, 115, 646 Linear manifold, 97 Linear matrix pencil, 288, 640 Linear operator, 40 nilpotent, 134 See also Linear transformation Linear ordinary difference equation: autonomous, 12 homogenous, 62, 63, 174 homogenous w/constant coefficients, 175
matrix, 495 nonhomogenous, 62, 63 periodic, 62, 63 Linear ordinary differential equation autonomous, 12 homogeneous, 12, 13, 62, 63, 138, 174 homogeneous w/constant coefficients, 148, 164, 175 matrix, 140, 495 nonhomogeneous, 12, 54, 62, 63, 145 periodic, 12, 13, 62, 63, 161 systems: continuation, 54 continuity with respect to parameters, 54 existence, 54 uniqueness, 54 See also Solutions of differential equations Linear space, 37 See also Vector space Linear subspace, 97,126 direct sum, 126 Linear system, 3, 60, 66, 94, 509 Linear transformation, 40, 100 bijective, 101, 102 determinant, 121 identity, 103 injective, 102 invariant, 127 nilpotent, 134 nullity, 101 orthogonal, 441 primary decomposition theorem, 133,134 principle of superposition, 66 projection, 126, 445 range space, 100 reduced, 127 representation by a matrix, 104 spectral theorem, 445 spectrum, 122 surjective, 101, 102 Linearization, 5, 48, 477, 504, 574 examples, 52 Linearized equation, 50, 51 Linearly dependent, 98, 234, 524 See also Vector Linearly independent, 97, 98, 99, 234, 524 See also Vector Lipschitz condition, 30, 45 Lipscitz constant, 30, 45 Lipschitz continuous, 30 LQG (linear quadratic Gaussian) problem, 352, 359 LQR (linear quadratic regulator) problem: continuous-time, 342 discrete-time, 348 LU decomposition, 647 Luenberger observer, 351 Lyapunov function, 468, 491 construction of, 474, 500 level curve, 471 Lyapunov matrix equation, 468, 469, 499 Lyapunov stability, 445, 492 linear systems, 452, 495 See also Stability Lyapunov's Direct Method, 468 Lyapunov's Second Method, 468
665 INDEX
666 INDEX
Magnetic ball suspension, 91, 92, 174-2 Mapping, 8 See also Function Markov parameter, 204, 387, 399 Matrix, 106 block diagonal, 128 characteristic polynomial, 123 cofactor, 113 column, 107 column rank, 119 companion, 461 companion form, 131, 153, 468 condition number, 646, 647, 649 conformal, 109 congruent, 436 controllability, 217 cyclic, 132 determinant, 111, 113 diagonal, 120 elementary unimodular, 526 eliminant, 542, 546 exponential, 57, 149, 150 fundamental, 139, 140 Gram, 234 Hamiltonian, 344, 349 Hankel, 399 Hermite form, 532 Hermitian, 648 highest degree coefficient, 526 Hurwitz, 460 idempotent, 126 identity, 110 ill-conditioned, 647 impulse response, 77, 166 indefinite, 443 inverse, 110 Jacobian, 48, 447 Jordan, 130, 135, 136 left inverse, 653 linear pencil, 288, 640 logarithm, 161 lower triangular, 131 LU decomposition, 647 minimal polynomial, 132, 133, 299 minor, 113,469 modal, 128 Moore-Penrose inverse, 652 negative definite, semidefinite, 443, 469 nilpotent, 134 nonsingular, 101, 110 norm, 43 null, 109 observability, 219, 253, 258 orthogonal, 441, 648 permutation of, 112 positive definite, semidefinite, 443 principal minor, 113, 469 proper rational, 391 properties, 107 QR decomposition, 647 rank, 101, 107, 306, 437, 524, 525 right inverse, 653 Rosenbrock, 301, 554, 564 self-adjoint, 443 singular, 110 square, 107 state transition, 56, 63, 143 Smith form, 531
Sylvester, 541 symmetric, 469, 648 symmetric part, 436 system, 301, 554, 564 system, extended, 557 Toeplitz, 259 transfer function, 78, 168, 178 transpose, 8, 107 unimodular, 526 unitary, 648 upper triangular, 131, 532 well conditioned, 647 Matrix difference equation, 496 Matrix differential equation, 140 Matrix fractional description, 517, 521 See also System representations Maximal element, 27 Zom's lemma, 26 McMillan degree, 397 Metric, 438 Metric space, 439 Microphone, 84 MIMO system, multi input-multi output, 66, 71 Minimal basis, 552 Minimal polynomial, 132, 133, 299 Mode of system, 156, 157, 186, 187 See also System Model matching problem, 374, 633, 644 Module, free, 552 Monic polynomial, 132, 298 Moore-Penrose pseudo inverse, 652 Motor, servomotor, 14, 206, 380 Natural basis, 100 See also Basis; Vector space Negative: definite, 437, 469, 491 indefinite, 469 semidefinite, 437, 469, 491 See also Function; Matrix Nilpotent operator, 134 See also Linear transformation Nonhomogeneous differential equations, 138, 145 See also differential equations Nonlinear systems, 3, 451, 477 Norm: Euclidean, 42 Frobenius, 653 induced, 43 linear space, 41,438 Manhattan, 42 matrix, 43 taxicab, 42 Numerical solutions of algebraic equations, 646 See also Solutions of algebraic equations Numerical solutions of differential equations, 25, 85, 86 See also Solutions of differential equations Numerical stability, 645 Observability, 169, 214, 219, 247, 249, 250, 253, 258, 263, 268, 274, 560 continuous-time system, 248, 252 discrete-time system, 219, 257
eigenvalue/eigenvector (PBH) test, 272 Gramian, 249, 253, 259 index, 295 matrix, 219, 253, 258 subspace, see Unobservable Observable, 562 eigenvalue, 268 mode, 268 Observer, Luenberger, 351 See also State observer Observer companion form, 203 Observer form, 292, 293, 296, 297 multi-output, 294 single-input, 293 Open covering, 18 Open loop control, 327 Operator, linear, 40,100 See also Linear transformation Optimal: control problem, 342, 348 estimation problem, 352, 357, 359, 362 Optimality principle, 371 Orbit, 492 Order of a realization, 394 Orthogonal, 439 basis, 440 complement, 441 linear transformation, 441 matrix, 441,648 projection, 445 Orthonormal basis, 439 Orthonormal set of vectors, 439 Output: decoupling zeros, 302, 564 zeroing, or blocking, property, 307 equation, 5, 58, 61 function controllability, 313 normal representations, 430 reachability, controllability, 312 vector, 4 Output feedback: dynamic, 363, 589, 642, 658 observer based, 363 static, 379 Parallelogram law, 438 Parseval's identity, 440 Partial state, 519 Partial sum, 21 Partially ordered set, 27 Peano-Baker series, 56, 143,147 Pendulum: inverted, 89, 380 simple, 16, 52, 82, 93 two-link, 83 Periodic solution, 492 See also Solutions of differential equations Periodic system, 11, 62, 161 Phase: plane, 198 portrait, 88 variable, 198 Picard iteration, 31 Piecewise continuous, 58 derivative, 23 function, 9 Plant, 182
Pole assignment problem, 330, 658 See also Eigenvalue or pole assignment Pole polynomial, 299 Poles at infinity, 319 Poles of a transfer function, 298, 299, 301, 488, 508, 563 Poles of the system, 298, 301, 564 See also Eigenvalue Polynomial: Hurwitz, 462 monic, 132, 298 stable, 462 Polynomial matrices, 524 column, row reduced, 526 common divisors, 535 coprime, left, right, 536, 612 division theorem, 545 doubly coprime, 593 equations, 540, 653 equivalent, 535 Hermite form, 532 proper or reduced row, column, 527 rank, 524 regular, 528 Smith form, 298, 531, 533 unimodular, 526 Polynomial matrix description, 517, 519, 521,553 See also System representations Polynomial matrix interpolation, 653 Polynomial of linear transformation, 104 Polynomial vector, degree, 526 linear independence, 524 Positive: definite, 437, 443, 469, 491 indefinite, 469 semidefinite, 437, 443, 469, 470, 491 See also Function; Matrix Positive innerwise, 498 Positive orbit, 492 Positively invariant, 492 Prediction estimator, 360 Proper rational matrix, 391 Proper transfer function, 170 Proper value, 102 See also Eigenvalue Proper vector, 122 See also Eigenvector Pythagorean Theorem, 439 QR decomposition, 647 Quadratic form, 436 Quantization, 183 Radially unbounded, 491 See also Lyapunov function Range space, 100 Rank, 107, 524 of a functional, 437 normal, 306, 525 test, 274 Rational function: biproper, 552, 577 proper and stable, 552
667
668 INDEX
Rational matrix: equations, 653 right inverse, 313 Smith-McMillan form, 299, 563 Rayleigh's dissipation function, 83 Reachability, 214, 215, 226, 235, 242,244 continuous-time system, 227, 235 discrete-time system, 215, 241 Gramian, 230, 236, 245 matrix, see Controllability output, 312 subspace, 228, 235, 242 See also Controllability, from the origin Reachable, 217, 228, 235 pairs, 228 state, 242 subspace, 217, 228, 235, 242 See also Controllable; Controllability, from the origin Real numbers, 8 Real sequence, 17 Real vector space, 38 Realization algorithms, 402 block companion form, 418 controller/observer form, 404 matrix A diagonal, 417 singular value decomposition, 423 Realization of systems, 383, 385, 388, 565 existence and minimality, 390, 394, 565 impulse response, 386 least order, irreducible, minimal order, 394,401,565 order of, 394 pulse response, 388 transfer function, 202, 388, 389 Reconstructible, 255 See also Constructibility Reduced order model, 424 Reference input, 326 Relation, 535 Response, 4, 59, 61 maps, 625 total, 165, 166, 174, 176 zero-input, 146, 165, 177 zero-state, 146, 165, 166, 177 Resultant, 542 Return difference matrix, 586 Riccati equation, 343, 349, 358, 363 continuous-time case, 342, 343, 358 discrete-time case, 349, 363 Ring, 552 Euclidean, 551 RLC circuit, 14, 16, 206, 276 Robust control, 327 Rosenbrock system matrix, 301, 554, 564 Rosenbrock system equivalence, 555 Routh-Hurwitz stability criterion, 466 Row, 107 Hermite form, 532 proper, see Polynomial matrices rank, 119 vector, 107 Row reduced, 526 See also Polynomial matrix Sampled data system, 65, 182, 183, 319 Sampling period, rate, 185
Scalar, 38 Schur-Cohn stability criterion, 498 Schwarz inequality, 42, 49,438 Semi-group property, 175 Sensitivity matrix, 599, 626 Separation principle, property, 325, 365 Sequence, 17 functions, 18 vectors, 44 Series, 203 Shift operator, 68, 92 Signal, digital, 183 Similarity transformation, 116, 120, 151 Singular, 101 value, 648 value decomposition, 423, 430, 648,650 vector, left, right, 651 See also Linear transformation; Matrix Skew adjoint, 443 Skew symmetric, 435, 436 Smith form, 298, 531, 533 See also Polynomial matrices Smith-McMillan form, 298, 299, 563 See also Rational matrix Solutions, bounded, 491, 493 Solutions of algebraic equations, 101, 115, 646, 655, 657 Solutions of difference equations: particular, 64 total, 64 Solutions of differential equations, 10, 12 bounded, 452 continuable, 27 continuation, 26, 27, 28, 31, 37, 45, 54 continuous dependence, 33, 37, 45 continuous dependence on initial conditions, 45,54 continuous dependence on parameters, 47,54 e-approximate, 23 Euler's method, 25, 85 existence, 25, 37, 45, 54 homogeneous, 58, 64, 145 noncontinuable, 27 particular, 58, 64, 145 Peano-Baker series, 56, 143, 147 periodic, 492 predictor-corrector method, 86 Runge-Kutta, 85 successive approximations, 31, 47,143 total, 58 uniqueness, 29, 30, 32, 45 See also Variation of constants formula Space: of linear transformations, 102 of ^-tuples, 38 of real-valued continuous functions, 39 span, 97 Spectral theorem, 445 Spectrum, 122 Spring, 16, 88, 320 Spring mass system, 13, 82, 83, 206, 320 Stability, 445, 492, 560 algebraic criteria, 461, 462 asymptotic, 159, 189, 449, 491, 494, 497, 504, 563
asymptotic in the large, 450, 491, 494, 499 attractive equilibrium, 449 bounded-input^ounded-output (BIBO), 481, 490, 505 causal, 67,70,71,77 domain of attraction, 449, 491 exponential, 449, 477, 480, 512 exponential in the large, 450, 459, 471 external, 481,490, 506 gap and position criterion, 464 global asymptotic, 491, 492, 494, 499 graphical, geometric criteria, 462 input-output, 452, 481, 490, 505 interlacing, 464 internal, 452, 481, 490, 613, 623 Leonhard-Mikhailov criterion, 463 linear systems, continuous, 159 linear systems, discrete, 189 Lyapunov, 445, 448, 590 Routh-Hurwitz criterion, 466 Schur-Cohn, 498 uniform, 448 uniform asymptotic, 449 uniform asymptotic in the large, 450, 459, 461 See also Equilibrium StabiUzable, 330 Stable, 159, 189, 448, 452, 455, 459, 491, 492 See also Stability Standard form: Kalman's canonical, 269 uncontrollable system, 264, 265 unobservable system, 267, 268 State, 56, 58, 60, 228, 235 controllable, 242 partial, 519 phase variable, 198 variables, 4, 56, 58, 61 vector, 4, 56, 58 State constructibihty, 251, 260 See also Constructibihty State controllability, 229, 235, 242 See also Controllability State estimator, 351,359 See also State observer State equation, 5, 58, 61, 165 linear solution, 55, 145 State feedback, 321, 322, 326, 605 input-output relations, 345 See also Feedback State observability, 250, 258 See also Observability State observer, 321, 322, 350, 359, 608 current, 360 full-order, 350, 358 identity, 350, 359 optimal, 357, 362 partial state, 354, 362, 608 reduced-order, 355, 362 State reachability, 215, 235, 242, 243 See also Reachability State space, 56, 58 State space description, 5 continuous-time, 58 discrete-time, 60 See also System representations State transition matrix, 56, 63, 143 State unconstructible, 250, 255
State unobservable, 248, 253, 258 Static: decoupling, 634 output feedback, 379 Structure theorem: controllable version, 291 observable version, 297 Subsequence, 18 Successive approximations, 31, 47 See also Solutions of differential equations Superposition principle, 60, 66 Sylvester: matrix, 541 rank inequality, 205, 396 resultant, 542 theorem, 437 Symmetric, 435, 436, 648 See also Bilinear functional; Matrix System: at rest, 71,77 autonomous, 12, 62 classification, 3 composite, 203 conservative, 81 continuous time, 4 description, external, 5, 65 discrete-time, 5, 60, 68, 174 distributed parameter, 3 dual, 222, 402 finite dimensional, 3, 4, 5 Fuhrmann equivalence, 554 Hamiltonian, 81 hybrid, 3, 183 infinite dimensional, 3 input, 66 inverse, 377, 634 linear, 60, 66, 94, 509 linear time-invariant, 59, 61 linear time-varying, 59, 61 lumped parameter, 3 matrix, 301,564 with memory, 67 memoryless, 66 mode, 156, 186, 187 multi input/multi output, 66, 71 nonanticipative, 67 nonlinear, 3, 451,477 output, 66 periodic, 12, 62, 161 realization, 383 response, 64 Rosenbrock equivalence, 555 sampled data, 65, 183 single input/single output, 66 strict equivalence, 555 time-invariant, 5, 242 time-varying, 5, 68 zeros, 301, 565 System interconnections: feedback, 203, 573 parallel, 203, 568 series, or tandem, 203, 570 System representations: balanced, 430 differential/difference operator, 517 equivalence, of, 170, 180, 554 external, 65
669 INDEX
670
fractional, left, right, 517, 521 input normal, 430 input-output, 66 integral, 72, 75, 76 internal, 5, 619 linear continuous-time, 76 linear discrete-time, 68 matrix fractional, 521, 612, 619 output normal, 430 polynomial matrix, 517, 519, 521, 553 standard form, uncontrollable, unobservable, 265,267, 268 state space, 5 zero-input equivalence, 170, 171, 173, 181 zero-state equivalent, 170, 171, 173, 181 Time reversibility, 145, 175 Toeplitz matrix, 259 Trajectory, 88, 198 Transfer function: McMillan degree, 397 pole polynomial, 299, 563 proper, 170, 530 strictly proper, 170, 530 Transfer function matrix, 78, 168, 177, 178 Transmission zero, 303, 564, 565 See also Zero Triangle inequality, 41 Trivial solution, 138,446 See also Equilibrium Truncation operator, 92 Two degrees of freedom controller, 622 Uncertainties, 327 Unconstructible, 250, 260 subspace, 250, 255, 260 See also Constructibility Uncontrollable, 264, 274 eigenvalues, modes, 265 See also Controllability Uniformly asymptotically stable, 449 Uniformly asymptotically stable in the large, 450, 459, 461 Uniformly BIBO-stable (boundedinput/bounded-output stable), 482, 483, 4t84 Uniformly bounded, 22 Uniformly continuous, 9 Uniformly controllable, 315 Uniformly convergent, 19 Uniformly stable, 448, 453, 454, 455, 459, 461, 470,511,512 Unimodular matrix, 526 Unit impulse, 74 Unit pulse, or unit sample, 68, 178 Unit pulse, or unit impulse, response, 70 Unit step: function, 154 sequence, 69
Unity feedback, 631, 642 Unobservable, 264, 274 eigenvalues, modes, 268 subspace, 220, 253, 258 See also Observability Unstable: complete instability, 273, 501 See also Equilibrium, unstable van der Pol equation, 16, 88 Variation of constants formula, 57, 146, 197 Vector, 38 coordinate representation, 100 linearly dependent, 97, 98, 234, 524 linearly independent, 98, 234, 524 normalized, 439 null, 38 unit, 439 Vector space, 37, 38, 97 basis, 99, 100, 170, 439, 552 complex, 38 dimension, 99 finite dimensional, 99, 437 inner product, 437 normed, 42, 438, 439 null, 100, 444 real, 38 Weierstrass M-test, 21 See also Infinite series Without memory, 66 Youla parameter, 595, 615, 635 Zero, 298, 563 blocking property, 307 decoupling input/output, 302, 564 direction, 306 at infinity, 320 invariant, 302, 303, 564, 565 polynomial, 302, 303, 564, 565 system, 301,564, 565 of transfer functions, 301, 302, 303, 564, 565 transmission, 303, 564, 565 Zero-input equivalence, 170, 171, 174, 181 See also System representations Zero-input response, 146, 165, 176 See also Response Zero-order hold, 183 See also Sampled data system Zero-state equivalence, 170, 171, 173, 181 See also System representations Zero-state response, 146, 165, 166, 177 See also Response Zom's lemma, 26 z-transform, 177